Home

Dokument_1.

Contents

1. Classification of dynamic nature of acceleration data After importing the raw acceleration data from a selected folder the dynamic nature is classified by the function dun natureClassification m see chapter 5 3 2 Training of the neural network For the supervised learning algorithm a training set with known output is necessary The allocation of activities to acceleration data is done by the user via the GUI To start this feature the function selectTrainData template m is called The allocated 41 5 IMPLEMENTATION training set is saved in the training folder and its features can be extracted This is computed by the functions described in chapter 5 3 3 and 5 3 4 Additionally the allo cated activity is saved in the feature vector One feature vector contains the features for a 6 second window and the index number of the activity in the activity catalog All the features from one training set are concatenated in one matrix It is called target matrix because it contains the target output of the classification These target matrices serve as input for the function trainNN m Function name selectTrainData_template m This function saves acceleration data with its actual activity information for training pur poses The absolute acceleration is plotted for visual representation of the signal A hard coded template is drawn consisting of several activities with specific lengths of time This template corresponds to the activ
2. A Dobra Decision Tree Classification in Encyclopedia of Database Systems New York Springer pp 765 769 2009 G Dorffner Klinische Signalverarbeitung und Mustererkennung Lec ture Medizinische Universtit t Wien Vienna 2013 K Deere A Sayers G Davey Smith J Rittweger and J H Tobias High impact activity is related to lean but not fat mass findings from a population based study in adolescents International Journal of Epide miology vol 41 no 4 pp 1124 1131 2012 P Duhamel and M Vetterli Fast Fourier transforms a tutorial review and a state of the art Signal processing vol 19 no 4 pp 259 299 1990 M Ermes J Parkka and L Cluitmans Advancing from offline to online activity recognition with wearable sensors in Engineering in Medicine and Biology Society 2008 EMBS 2008 30th Annual International Con ference of the IEEE pp 4451 4454 2008 D Figo P C Diniz D R Ferreira and J M Cardoso Preprocessing techniques for context recognition from accelerometer data Personal and Ubiquitous Computing vol 14 no 7 pp 645 662 2010 XII FNRA10 FrJo98 GCDC10a GCDC10b GCDC13 GeMCO06 GjGC10 HeKi08 HGP 11 HrLe05 REFERENCES K Frank M J V Nadales P Robertson and M Angermann Reliable real time recognition of motion related human activities using MEMS inertial sensors in Proceedings of t
3. Installation of MovelT The software code is saved on the DLR internal server It is saved in several MATLAB compatible m files 1 Goon the DLR internal server to G WP IT MovelT 2 Copy the folder Program Code to your local hard drive Starting MovelT To run the MovelT program the code needs to be executed from MATLAB 1 Start MATLAB 2 Press Open go to the folder where the MovelT software is saved and choose the file MovelT_main m el Open 3 On the tab EDITOR press on the green arrow saying Run ou WEEN BsA se Fy D EE en Breakpoints Run F eile Find y EDIT NAVIGATE BREAKPOINTS 4 If this is the first time you run the MovelT software the following window will pop up Press Add to Path to continue MATLAB Editor ez i File D jos_de NutriHEP HAPR Software HAPR_gui m is not found in 2 the current folder or on the MATLAB path To run this file you can either change the MATLAB current folder or add its folder to the MATLAB path Change Folder AddtoPath Cancel Help 5 The MovelT software will start showing the home menu For more information see Getting started with MovelT on page xiii APPENDIX 2 2 Working on a computer without MATLABO license If you do not have a valid MATLABO license on your computer you need to install the MATLABO Compiler Runtime and the MovelT software Please note Without a MATLAB license it is only possible to work
4. fore this function is called calcKhan_movWin m and the features are referred to as Khan Features in the software code First the data is divided into 50 overlapping moving windows Furthermore the SE on all three axes is calculated by a sub function calc SE m The AR coefficients are estimated by a built in MATLAB function using Burg s method For the SMA the sums of the absolute accelerations on each axis are summed up see chapter 3 2 3 Calculate Khan features for every window The SE is calculated for every axis by the sub function calc_SE m This function cal culates the SE of the acceleration signal for a certain frequency band The calculation is done according to KHAN ET AL KLLK10 as described in chapter 3 2 2 After an FFT the PSD is calculated SE is calculated for each window as in Equation 3 9 in chapter 3 2 2 The FFT is calculated as described before For the estimation of the PSD the built in function dspdata psd is used This function generates the PSD values and 39 5 IMPLEMENTATION the sum at all frequencies can be calculated The division of this sum by the logarithm of the number of frequency components in the corresponding band gives the SE spectral entropy SE SEX calc SE winX rate SEY calc SE winY rate SEZ calc SE winZ rate The AR model is estimated with Burg s method and the coefficients are computed by the built in arburg function According to KHAN ET AL KhLK
5. software The free run simulates more practical conditions and can therefore be re garded as more representative and its evaluation results as more suitable However the free run performance parameters showed a higher SD and more outliers A free run without any protocol allowed the subjects to do any kind of locomotion Some subjects wanted to put the software on trial by moving in unusual ways e g climbing stairs backwards or doing sit ups This led to a higher number of unclassifiable activities in relation to the total acquisition which led to a lower accuracy and reliability The classifying system in this thesis has a similar structure like the classifier by LEE ET AL LKK 10 and its accuracy parameters are similar as well However neither LEE ET AL nor other authors published any statement about reliability of a classifier 65 7 DISCUSSION AND CONCLUSION The classifier developed in this thesis is able to correctly predict new and unknown data with a sufficient accuracy and reliability The speed of the classification is decent Training for each subject is needed beforehand but the resulting ANN is applicable for the rest of the study An approximate computation time of 6 minutes per measurement day is considered adequate because there is no mention of any minimum computation time in literature The compliance check showed reasonable results but its accuracy needs to be proven in a further evaluation During the work on this t
6. stairs down stairs up walking slow walking normal res lying sitting standing unclassifiable Activity class time MM SS Figure 6 5 Classification result example of a 2 run The acceleration signal is plotted in the upper part Below the results of activity classi fication are shown Every blue dot stands for a classification of the activity according to the axis of ordinates Unsuccessful classifications are defined as unclassifiable Ac cording to Table 6 2 it is known which activity was performed by the subject True ac tivity and classification result are comparable Therefore the numbers of correct and incorrect classifications can be obtained To evaluate a classifying ANN system a confusion matrix was created How to fill out a confusion matrix and how to quantify performance parameters was described in chapter 3 5 The numbers of successful and unsuccessful classifications were ob tained by counting the blue dots in the classification results graphic see Figure 6 5 The creation of the confusion matrix for an AC classification was an unambigu ous task When and which movements were performed was clearly specified by the order of the activity course see Table 6 2 In contrast interstation activities were ignored Evaluating the FR classification results was more difficult Type and order of movements were not specified in the third run but instead had to be interpreted b
7. threshold EE H gr Stat re activity else dyninatuce 277519 es dynamic activity end take timestamp in next to the middle of window dyn nature 1 1 dataTime winStart halfWin 1 37 5 IMPLEMENTATION 5 3 3 Feature extraction time domain analysis Function name calcTime_movWin m This function extracts features of the time domain for classification purposes First the data is divided into 50 overlapping moving windows For every window mean SD and absolute maximum of the data is calculated Calculate features for every window All features in the time domain can be calculated with the built in functions mean std and max These three features are assembled in a feature matrix where every row con tains the features of one moving window This is applied to every axis of the accelera tion data for i 1 N calc mean of one axis meanValue mean winData calc standard deviation of one axis stdValue std winData calcabsolut maximum acceleration in one axis maxValue max abs winData o assemble features as array featMatrix i meanValue stdValue maxValue end 5 3 4 Feature extraction frequency domain analysis Function name calcFFT_movWin m This function applies an FFT on acceleration data and extracts frequency domain fea tures for classification purposes First the data is divided into 50 overlapping moving windows For every window an FFT is performed The power of the
8. 2 Order of activity Course aci 58 Table 6 3 Accuracies and reliability of classifier mean SD nn 62 Table 6 4 AC confusion matrix from bp 63 Table 6 5 AC confusion matrix from tibia oooooccccnnnccccnnccanaccnonnnonnnnnanrrnncnnnnnnnnnnnnnn nn no 64 Table 7 1 Requirements and their accompltsbmente nono 67 LIST OF ABBREVIATIONS List of abbreviations DO Activity course PIN E Artificial neural network AR an Autoregressive Gina Compact disc A tention Comma separated values DAT osa Data file DE Tuna Discrete Fourier Transform DEN vasca Deutsches Zentrum f r Luft und Raumfahrt German Aerospace Center Fiese Fast Fourier Transform FFTW ee Fast Fourier Transform in the West Fries Free run EE RER Gulf Coast Data Concepts EE E Graphical user interface HEP ae HEPHAISTOS study Dissen Identification MOB MATLAB Compiler Runtime MEM Standar Micro electromechanical system NASA National Aeronautics and Space Administration de E Nutrition study using HEP orthosis Ft ras Power spectral density FOO DEE Requirement Oya Standard deviation Espadas Spectral entropy MA Signal magnitude area SysReQ System requirement PA Text file USB Universal Serial Bus 1 INTRODUCTION 1 Introduction 1 1 Motivation With respect to future long duration space missions to Mars or beyond muscle and bone loss are a major concern at the astronaut s health Since these processes h
9. 2 The program explains itself totally partially partially totally disagree disagree agree agree O0 O o 2 O 3 The program will save me time in my work Very unlikely Fairly unlikely Maybe Fairly likely Very likely Oo O Oo 2 Diagrams 1 The diagrams give a good overview of the data totally partially I can t partially totally disagree disagree decide agree agree e LI LI O 2 The diagrams are simple totally partially I can t partially totally disagree disagree decide agree agree L Page 3 of 4 APPENDIX 3 The diagrams explain themself totally partially I can t partially totally disagree disagree decide agree agree e E CI O CI L 4 would use a diagram in my future work Very unlikely Fairly unlikely Fairly likely Very likely LI 5 Fairly likely Very likely User manual 1 The user manual is helpful Very unlikely Fairly unlikely Maybe Fairly likely Very likely e LI CI LI CI 2 The user manual is understandable Very unlikely Fairly unlikely Maybe Fairly likely Very likely E LI LI LI LI 3 Do you feel confident to work with MovelT only with manual instructions Very insecure Fairly insecure can t decide Fairly confident Very confident e E L L LI LI Page 4 of 4 APPENDIX 9 2 User manual The following pages contain the user manual for the developed s
10. 2013 08 26 10 16 cations The lower graph 100 contains the corresponding 80 certainty levels of the classi fications time MM SS 60 Certainty 40 time MM SS xvii APPENDIX Activity distribution of Tibia 3 3 7 41 BEE unclassifiable EEE standing BEE sitting FF tying walking normal i walking slow MES stairs up BEE stairs down EEE jogging In the middle of the screen the activity distribution is represented in a pie chart Per centages refer to the total time of data acquired Days Hours Minutes Seconds elapsed time 00 00 14 59 565 Avg certainty of Classification dynamic activities 99 7922 static activities 95 7240 all activities 97 5903 No of recognitions Activity distribution estimated Time dd hh mmiss unclassifiable 10 3 3003 00 00 00 29 standing 125 41 2541 00 00 06 11 sitting 6 1 9802 00 00 00 17 lying 8 2 6403 00 00 00 23 walking normal 88 29 0429 00 00 04 21 walking slow 9 2 9703 00 00 00 26 stairs up 9 2409 00 00 01 23 stairs down 6 9307 00 00 01 02 jogging 2 6403 00 00 00 23 sprinting 0 00 00 00 00 SUM 100 00 00 14 59 On the right of the screen the distribution is shown again in absolute numbers orga nized in a table It also contains the total time and the average certainty level of classi fication xviii APPENDIX 5 4 How to check for
11. 68 A etea ae anna EE ENa A ea EAEE EAE AE EAEE EE A ERRER xl ADPENAR ee i 9 1 Usability test questionnaire un i 9 24 Userman al EE vi VII LIST OF FIGURES List of figures Figure 2 1 Example of graphical visualization for Hei 3 Figure 2 2 Example of graphical visualization for He 4 Figure 3 1 Artificial neural network topology ENNEN 16 Figure 3 2 USB Accelerometer X6 2 GCDC Waveland USA 18 Figure 4 1 Training and classification Concept 25 Figure 4 2 Data flow dlagram een 26 Figure 5 1 Screen for allocating during training phase with activity template 43 Figure 5 2 Classification result GU 48 Figure 5 3 Pie chart results GUl u n uenr une 49 Figure 5 4 Compliance results GUI tota he aka 51 Figure 6 1 Position of accelerometer on HEPHAISTOS orthosis a and 9n Shin Quard Da teas e a A E cies 53 Figure 6 2 Belt bag with accelerometer Au 54 Figure 6 3 Subjects wearing tibia and hip sensor a GoPro chest mount harness b and GoPro camera point Of View cl 55 Figure 6 4 Map of DLR campus with building 24 and locations of the stations 57 Figure 6 5 Classification result example of a2 run 60 VIII LIST OF TABLES List of tables Table 3 1 Converting rules from raw counts data into g units GACDC10al 18 Table 3 2 Confusion matrix exvample ENNEN 19 Table 4 10 Feature Mt aa a len 30 Table 5 1 Table results GUI a ee rg 50 TADIG ed 56 Table 6
12. L Wagstaff Intelligent clustering with instance level constraints Cornell University Ithaca NY 2002 T Weber M Ducos E Mulder F Herrera G P Br ggemann W Bloch and J Rittweger The specific role of gravitational accelerations for arterial adaptations Journal of Applied Physiology vol 114 no 3 pp 387 393 2013 T Weber M Ducos P Yang D Jos P Frings Meuthen G P Br ggemann W Bloch and J Rittweger The HEPHAISTOS study Compliance and adherence with a novel orthotic device for calf muscle unloading Journal of Musculoskeletal and Neuronal Interactions to be published 2013 S Weber Carstens J Schneider T Wollersheim A Assmann J Bier brauer A Marg H Al Hasani A Chadt K Wenzel and S Koch Criti cal Illness Myopathy and GLUT4 Significance of Insulin and Muscle Contraction American Journal of Respiratory and Critical Care Medi cine vol 187 no 4 pp 387 396 2013 C C Yang and Y L Hsu A Review of Accelerometry Based Wearable Motion Detectors for Physical Activity Monitoring Sensors vol 10 no 8 pp 7772 7788 2010 XVII 9 Appendix Usability test questionnaire User manual APPENDIX APPENDIX 9 1 Usability test questionnaire Questionnaire as part of the usability test filled out by conducting researcher of the NutriHEP study 1 Usability test of MovelT 1 1 Interview Please answer the followin
13. Subsequently the hip data classification is performed All results are plotted and saved in a separated classi fication folder Function name activityClassification m This function classifies accelerometry data into different activity classes After the im port of data and its time shift application the data is classified by its dynamic nature Following this static data is classified with a static ANN and dynamic data with a dy namic ANN Finally all results from the classification are saved in separated matrix files for further access and applications Import acceleration data and classify dynamic nature Acceleration data is imported with the multiple RAW Conversion m function see chapter 5 3 1 If a time shift file is present it is applied to one of the sensors The TXT file contains a shift in milliseconds which was calculated during time synchronization see chapter 5 2 2 It is simply added to every time stamp of the acceleration data of the regarding sensor The dynamic nature of the signal is analyzed by calling the func tion dyn_natureClassification m see chapter 5 3 2 Classify static and dynamic activities separately All static data is concatenated and the static features are extracted as described in chapter 5 3 3 The resulting static feature matrix is saved as backup and then fed to the static ANN The same process is executed with the dynamic data 45 5 IMPLEMENTATION The classification results a
14. a course had to be completed which required the follow ing activities Standing Sitting Lying Walking with normal pace Walking with slow pace Climbing ascending stairs Climbing descending stairs Jogging with slow pace Table 6 1 Activity classes Static activities Dynamic activities Standing Walking with normal pace Sitting Walking with slow pace Lying Climbing ascending stairs Climbing descending stairs Jogging with slow pace Subject stands in upright position Both legs are equally loaded so that the hip remains in its horizontal position Other move ments are to be avoided Subject sits on a chair Both feet touch the ground Other movements are to be avoided Subject lies on its back on a couch or table Other movements are to be avoided Subject walks across a corridor This will be done in a uniform everyday pace so a 50 meter range is completed in approximate ly 30 seconds Subject walks across physiology lab This will be done in a uni form slow pace so a 20 meter range is completed in approxi mately 30 seconds Subject climbs staircase several floors up The movement should be uniform and slow paced No skipping steps or running al lowed Subject climbs staircase several floors down The movement should be uniform and slow paced Overloading or shock absorb ing steps are to be avoided Subject jogs in a slow pace across a corridor The pace should be unif
15. a pool of features which are based on unsupervised learning Prus06 Hidden Markov model A hidden Markov model is a generative probabilistic model for analyzing time series data At each time step it consists of a hidden variable and an observable variable This Markov process is based on two assumptions First a hidden variable depends only on the previous hidden variable Second an observable variable depends only on the hidden variable at that time step Estimations of probabilities between observable and hidden variable allow a prediction and classification KNEK08 Conditional random field A conditional random field is a discriminative probabilistic model and has the same structure as the hidden Markov model Instead of directed dependencies between vari ables the conditional random field is an undirected graphical model Conditional prob abilities have been replaced by corresponding potentials Estimating these potentials allows a classification KNEK08 Artificial neural networks The human brain has one of the most complex structures found in nature Its ability to perform cognitive tasks is superior to modern computers thanks to numerous neurons working parallel The human brain consists of approximately 100 billion 10 neurons which are connected with one another by 10 to 10 synapses Stimulated neurons pass the signal on to other neurons The signal of the stimulus can be amplified or re duced by the synapses Hebbi
16. a signal by fewer values Consequently a feature extraction of the acceleration data is needed to make it identifiable and classifiable Numerous methods of feature extraction are available Algorithms derived from de scriptive statistics HGP 11 DSD 12 GeMC06 and the Fourier analysis HGP 11 KLLK10 are mostly used Further explanation of these methods will be given in chapter 3 2 The raw signal data from the accelerometers are not scaled in units of accelera tion but in raw count data Moreover the signal data may contain noise or other inter ferences To this end the raw data has to run through another procedure before fea ture extraction The conversion into units of acceleration is dependent on the accel erometer settings Several filtering methods are used for noise reduction moving aver age KLLK10 low pass filter HGP 11 GeMX06 LoYA09 and band pass filter YaHs10 Further explanation of these methods will be given in chapter 3 1 An evaluation of the classifying system is necessary to verify the accuracy of the software program Since no requirement for minimum accuracy was formulated a literature research had to be performed Research showed that classification accuracy varies from 80 KNEK08 LoYA09 LKK 10 to over 90 PGKHO9 KwWM11 KLLK10 KhLK08 HGP 11 More detailed information of other literature will be cov ered in chapter 3 6 The required user s accuracy is thus set to 80 for the software
17. at the cor responding frequency axis value FDFC10 It has been shown that different physical activities contain different dominant main frequencies HGP 1 1 PGKHO9 FCFD10 Fast Fourier Transform The Fourier Transform is the link between representing a signal in its time domain or in its frequency domain In the time domain the signal is represented with a time depen dency the frequency domain shows a frequency dependency Physical signals are often acquired in a time discrete manner resulting in equidistant sampled data The Discrete Fourier Transform DFT is applicable for the numerical analysis of such sig nals M ll13 10 3 MATERIALS AND METHODS NEUBAUER Neub12 explains that the DFT allocates to a finite signal series in the time domain of length N with index0 lt t lt N 1 x t oeren 1 x 0 x 1 xCN D Equation 3 1 adapted from Neub1 2 the finite spectral series in the frequency domain XP osrew 1 X 0 RO X N D Equation 3 2 adapted from Neub1 2 consisting of N spectral values with index 0 lt f lt N 1 The transformation rule is N 1 Equation 3 3 XP A ster j er adapted from Neub12 The DFT is written as X f DFT x t Equation 3 4 adapted from Neub1 2 The FFT is another way to calculate the Discrete Fourier transform It approaches the same results like other methods but is more efficient and less time consuming It can reduce the computation time of an input
18. attached to a subject s chest three kinds of activities resting walking and running were classified SMA and medi an frequency were the only two features used as input for a decision tree The ground truth data for evaluation purposes was recorded by a following annotating researcher An accuracy of 81 25 was achieved The focus of LONG ET AL LoYAO9 was the comparison of a Bayes classifier and a decision tree classifier in assessing daily energy expenditure In this work acti vities like walking running cycling and driving were classified In a naturalistic envi ronment without researchers acceleration data was acquired with a single tri axial ac celerometer placed on the waist Annotations of activities for a ground truth data set were written down by subjects The allocating of acceleration data to activities was done by hand SD periodicity and orientation features were input for a decision tree classifier and Bayes classifier Accuracies of 72 8 and 71 5 respectively were achieved KHAN ET AL KLLK10 were able to identify new features which allow activity recognition independent of the sensors position The intention was to establish a recognition system for an unsupervised free living environment A single tri axial accel erometer was worn either in a chest pocket in a trousers pocket front or rear or in an inner jacket pocket Classified activities include resting walking climbing stairs run ning cycling and
19. data SaWe11 19 3 MATERIALS AND METHODS For every column the reliability can be calculated as the number of correct classifications green diagonal element divided by the total number of classifications for this activity sum of column This leads to a reliability of 66 7 for the activity Walking meaning that two third of all Walking classifications actually represent Walking in the input data Accuracy and reliability can be calculated for every activity and are written in yellow cells in Table 3 2 The averages of these values are performance pa rameters of the classifier Average accuracy and average reliability are highlighted in grey cells A third parameter is the overall accuracy It is calculated as the total number of correct classi fied activities diagonal elements divided by the total number of input activities num ber of classifications ITC113 Usability questionnaire A new GUI has to be developed for the custom software To evaluate the usability of the software and its user interface a full usability test would be appropriate However such a test in all its dimensions is too time consuming and beyond the scope of this thesis Therefore a rough quantification of the ease of use is accomplished by a short usability questionnaire The conducting researcher of the NutriHEP study is asked to read the user manual See appendix 9 2 and to work on the software If the user manual is not help ful for unpredictable
20. description SysReq3 Signal processing methods for pre processing SysReq4 Evaluation method for determining user s accuracy SysReg5 Activity classification with more than 80 user s accuracy SysReq6 Numerical computing environment for programming SysReq7 Decision rule for checking compliance SysReq8 User friendly graphical interface SysReq9 Compiling software for generating standalone executables 3 MATERIALS AND METHODS 3 Materials and methods Analysis of acceleration data needs various procedures and methods First the raw data must obtain a pre processing Subsequently the resulting signal is characterized by certain features Based on these combined feature patterns different activities can be identified by a classification method Finally the classifying system is tested in an evaluation According to these required procedures several methods for pre processing feature extraction classification and evaluation are under consideration This chapter explains possible methods for further analysis 3 1 Pre processing methods Filtering Digital filters are used for separating signals that have been combined and for restoring distorted signals Analog filters are applicable for the same tasks but digital filters achieve better results Smit03 Moving average filters as they are used in KLLK10 are mostly incorporated to filter out random noise Each point in the output signal is produced by averaging a number of poin
21. direction turn accelerometers ON stand still and wait for climbing stairs 30s Staircase 0 30 Climbing ascending stairs walk 4 floors upstairs keep uniform pace at 3 floor stop and stand still 90s Interstation activity stand still for 15 seconds turn around stand still and wait for 2 30 30s 2 30 Climbing descending stairs walk 4 floors downstairs keep uniform pace no skipping or jumping at end of stairs stop and stand still 90s Base ment Interstation activity stand still for 15 seconds walk to end of tunnel and turn around stand still and wait for 5 00 60s 5 00 Walking with normal pace walk on own pace go around corner until black door stop and stand still 90s Interstation activity stand still for 15 seconds turn around stand still and wait for 7 00 30s 7 00 Walking with normal pace walk back on own pace to end of tunnel stop and stand still 90s Interstation activity stand still for 15 seconds walk to running start stand still and wait for 9 30 60s 9 30 Running as jogging jog to end of corridor go around corner until black door stop and stand still 30s Interstation activity stand still for 15 seconds walk to ground floor stand still and wait for 11 30 90s Ground floor 11 30 Walking with slow pace walk slowly acro
22. elements were in the unclassifiable column This means that the classification certainty was not high enough but it only affected the accuracy of the classifier However the most acti vity activity misclassifications were between Sitting and Standing This indicates diffi culties for the classifier distinguishing these two activities Looking at the accuracies on the right column the lowest producer s accuracy was measured for the activity Walking slow A higher reliability of 89 2 indicates that the classifier is able to distinguish such locomotion from others The remaining reliabilities are over 80 The same descriptions are applicable to the results of the tibia sensor during the AC classification in Table 6 5 With both feet on the ground while sitting the orien tation does not change during standing This explains the lower accuracies for Sitting and Standing Again the accuracy for the activity Walking slow is the lowest Activity course classification with hip and tibia sensor is accurate and reliable for almost every of the activities stated in Table 6 1 The activities Standing and Sitting are more often confused than others Half of the activity Walking slow is not identified but nearly 90 of Walking slow classifications is correct Such a confusion matrix as in Table 6 4 is not suitable for the free run classifi cation Time spent per activity varies between subjects because every run was chosen 61 6 EVALUATION
23. function checks for a TXT file named config txt It contains information about the subject and the body position the accelerometer was worn How this information is 32 5 IMPLEMENTATION changed in this file is explained in the user manual see appendix 9 2 If no configura tion file is found the software throws an error saying that a wrong drive was chosen Starting time and date of the recorded data is read out through the data files them selves Every data file contains a header with such information This function is exclu sively adapted to GCDC accelerometers Import files from USB Drive After checking for a valid saving path the number of files for import is gathered A gen erated progress bar showing the computing progress according to the number of im ported files is shown on the screen For every file date and time of acquisition are read from the header According to the date the calendar week number is calculated with an external function If the corresponding folder is not existent it will be generat ed After that the data is renamed and copied to the right target folder Any error in the process described above i e missing header information will be noted in a list of failed imported data This list will be shown after trying to copy every data file from the cho sen accelerometer An abortion of the saving process is possible by closing the loading progress bar window 5 2 2 Time synchronization Fu
24. how long the subject moved in a certain way According to the interviewee a representation in the form of a pie chart would be helpful because it provides a proper overview of the relative distribution of the different activities Reg Was the orthosis worn the entire time The analysis software should have the ability to tell when the orthosis was worn or not worn by the subject Figure 2 2 shows an example of a weekly overview which was created in collaboration with the conducting researcher In this graphic the compliance is binary color coded It is easy to detect time intervals where the orthosis was not worn for a quick interrogation with the subject Furthermore this requirement implies that the software should be applicable with the usage of an orthosis Subject A Orthosis not worn Monday Tuesday Wednesday Thursday Friday Saturday CW 20 CW 21 orthosis worn orthosis not worn Figure 2 2 Example of graphical visualization for Req3 Req4 What does the data look like during unidentified activities The analysis software should have the ability to show raw acceleration data collected during unidentified activities for further manual investigation Req5 The software should be easy to use Interviewing the conducting researcher of the NutriHEP study it turned out that the computational skills of potential users may bear insecurities It was stressed that the handling of the software shou
25. known classes or activities are used as input because their results are required to fill out a confusion matrix Table 3 2 presents an example of a confusion matrix for a three activity classifier to differentiate between Standing Walking and Running Table 3 2 Confusion matrix example Classification result Confusion matrix Standing Walking Running Accuracy Standing 10 0 0 100 0 True activity Walking 0 6 1 85 7 Running 0 3 8 72 7 bici ccuracy Reliability 100 0 66 7 88 9 86 15 Average Reliability 85 19 85 71 EE ccuracy Number of classifications 28 The second row indicates that 7 data inputs were of the activity Walking Six of those were classified correctly one was misclassified as Running The numbers of correct classifications are highlighted in green on the diagonal of the matrix Off diagonal ele ments contain misclassifications and are colored in red The sum of a row is the total number of inputs of one activity The sum of a column is the total number of classifica tions of one activity class For every row the accuracy can be calculated as the number of correct classifi cations green diagonal element divided by the total number of inputs for this activity sum of row This leads to an accuracy of 85 7 for the activity Walking Accuracy is a measure of the degree to which the predicted activity of the classifier matches the true input
26. lyzing daily physical activity of human beings By using MovelT it is possible to adapt classifying systems to an individual person to view activity profiles or raw accelerome try data Its use requires the application of USB accelerometry sensors by Golf Coast Data Concepts This manual shows how to install the MovelT software how to configure the acceler ometry sensors and how to start the analyzing procedures Minimum system requirements e Golf Coast Data Concepts accelerometers Recommended system requirements e Golf Coast Data Concepts accelerometers e Mathworks MATLAB R2012b o Neural Network Toolbox o Signal Processing Toolbox Formatting definitions Italics Brand name Bold Folder or feature button name Underlined Link to chapter in this manual Underlined blue Link to server folder or website 2 Installation The Movement Identification Tool MovelT software is based on and therefore only works with Mathworks MATLAB Before working with MovelT you need to find out if the computer you want to work on has a valid MATLAB license installed Additionally you need the Neural Network Toolbox and the Signal Processing Toolbox Any version since MATLAB R2012b is recommended Ask your departmental IT administrator for further help APPENDIX 2 1 Working on a computer with MATLAB license If you have access to a valid MATLAB license and the essential toolboxes you only need to install the MovelT software
27. menu Import accelerometer raw data All CSV files are imported at once with the built in dir function From the first one its starting time sampling rate and gain are read out from the header information These details are important for the conversion process and higher level functions Convert raw data to g forces According to the gain setting the correct calibration is chosen from the user manual see Table 3 1 on page 18 For a high gain setting which is used to acquire accelera tions between 2 g and 2 g the raw count data is divided by 1024 For a low gain set ting allowing acquisition in a range from 6 g to 6 g the raw count data is divided by 340 These calculation specifications are taken from the user manuals of X6 7A and X6 2 accelerometers by GCDC GCDC10a The time stamps saved in the CSV files are also retrieved and saved as time array In the accelerometers time is saved as elapsed seconds since the start of the measurement Therefore the relative time array 35 5 IMPLEMENTATION needs to be converted into an array with absolute serial date numbers This is accom plished by adding the start time to every time stamp with the built in addtodate func tion Again the seconds need to be converted into milliseconds o Read out time array rawT raw data 1 convert relative time array to absolute serial date number array dataTime zeros length rawT 1 for i 1 length rawT dataTime i
28. of the NutriHEP study throughout the whole thesis Further would like to thank Prof Dr Oliver Kalthoff for his supervision Thanks are also due to my colleagues at the Department of Space Physiology who volunteered as subjects during the evaluation study also thank Michel Ducos and Prof Dr J rn Rittweger for their constructive comments in the early stages of this work Valuable advice in scripting this thesis is gratefully acknowledged to Josefine Kupper and Miriam Reisenhofer LIST OF CONTENTS List of contents Abstrac ENglish EE Il ee A II Eidesstattliche Erkl rung WEE IV Ack owledgement eccna ec V Eo ESEE e ee VII re IX MS O x Tin ele EEN aan 1 Ted e le ke 1 1 2 Tesis Outline re 2 2 e EE 3 2 1 User reg iremeniS EE 3 2 2 Sysiem reqliremenist 2 aA ie 6 3 Materials een une EE 8 Sela Preprocessing Methods una ea en 8 3 2 Feature extraction methods use 9 3 2 1 Descriptive statistics uuuun444nnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn 9 3 2 2 A Ee eee ean noes Maree 10 3 2 3 Miscellaneous fealures nu nee 13 E DE ee e e n 14 34x e 17 3 5 Evaluation methods nee eek denied 19 3 6 LALO OT ING ar une 21 Ais Design CONCERN reacts ee elements 23 4 1 Classification with artificial neural networks ooononnnonoccccnnnncccnnccanccccnnnccnnnnnnns 23 4 2 Dala tow POCOS Sie irte 24 AD Data acquIStION ne mas ee 27 4 2 2 Dynamic nature classification nen nnnnnnnenn
29. some subjects were excited during the first run because the course was new and unknown It cannot be ruled out that the subject paid more attention to its movements under the experimental conditions Additionally the second run seemed to be boring because of its redundancy The subject might have paid less attention and moved in a more ordinary way To eliminate these possible effects a randomization was applied whether the first or the second run served as training set Further the position of the tibia sensor was randomized Five subjects wore the shin guard on the left leg the rest on the right leg For a better differentiation the classification of the first or second run is referred to as activity course AC classification The classification of the third run is referred to as free run FR classification Evaluation results The results of a classification were compared to the actual activities performed by the subject For the second run the order was given in Table 6 2 For the third run the results had to be compared to the video recording An example of a classification result is shown in Figure 6 5 59 6 EVALUATION Absolute acceleration of Tibia starting on 2013 08 26 10 16 Dm d acceleration g time MM SS Activity classification of Tibia starting on 2013 08 26 10 16 sprinting r _ Jogging
30. sort dynamic nature tibiaTimes CCtibia winTimesDyn CCtibia winTimesStat ones 1 length CCtibia winTimesDyn zeros 1 length CCtibia winTimesStat hipTimes CChip winTimesDyn CChip winTimesStat ones 1 length CChip winTimesDyn zeros 1 length CChip winTimesStat tibiaTimes sortrows tibiaTimes 1 hipTimes sortrows hipTimes 1 46 5 IMPLEMENTATION Compliance check The compliance check looks at every sample but 120 samples before and after every sample are taken into account By this only longer episodes of non compliance are recognized If for all samples the tibia signal is classified as static and the hip signal as dynamic then the compliance is classified as not present i e the orthosis is not worn Compliance is binary encoded as 0 orthosis not worn and 1 orthosis worn w 120 if tibia is not dynamic and hip is dynamic then tibia sensors is not worn for i w 1 N w if tibiaTimes 2 i w itw hipTimes 2 i w itw compliance 2 i w itw 0 end end The function plotCompl m shows a graphical visualization of the compliance check results The interfaces of all plotting functions are described in the following chapter 5 5 Graphical visualization of results 5 5 1 Classification results Acceleration data The end result of a classification may look like the example GUI in Figure 5 2 The ac celeration signal is plotted over time on the
31. subject s compliance MovelT provides the feature to check for the subject s compliance by verifying if hip and tibia sensor were worn simultaneously 1 Press the button Compliance 2 Browse to subject folder and choose folder containing Hip Tibia and Classifi cation folder Pop up window shows the results of the compliance check 4 Its graphical representation is saved as FIG file in the folder chosen in step 2 gt Graphical representation Compliance check of Tibia sensor starting on 2013 08 15 14 52 00 00 time MM SS BEE worn MA not worn The compliance is represented in a colored timeline where green and red stand for timeframes with worn and not worn tibia sensor 6 Working with a single sensor MovelT is also designed for working with a single accelerometry sensor for usage out side of the NutriHEP study The training und classification procedures with a single sensor correspond to the instructions above in 5 2 How to train a neural network on page xvi and 5 3 How to classify activities on page xvii The training feature just needs a naming of the sensor s position beforehand This naming is important since the soft ware looks for the data in a folder of the same name The steps of time synchronization and compliance check are not applicable xix 7 Attachments APPENDIX 7 1 Default training course Starting WE i Posten ve Gace CI E Base ment 1 0 00 Standing face walking
32. template It is not obligatory to use the template for activity allocation The template is supposed to improve the time efficiency concerning the application during the NutriHEP study Single activities can be allocated by hand Further types of locomotion like cy cling or sprinting are not covered by the template because these movements are pro hibited during the intervention study However new activities can be added to the acti vity catalog When naming new activities the dynamic nature needs to be specified If the new activity is of static nature the number of static activities in the activity catalog Nstatic is incremented Nstatic is important for the assigning of activities to the right sub folder 43 5 IMPLEMENTATION Function name trainNN m This function trains an ANN by solving a pattern recognition problem This code lines are mainly generated with the assistance of the Neural Network Pattern Recognition Tool nprtool It is part of the Neural Network Toolbox in MATLAB Create a pattern recognition network The settings of the ANNs basically are according to HANSON ET AL HGP 11 The hid den layer has a size of 20 neurons HANSON ET AL stated that a difference of 5 neu rons did not affect the results hiddenLayerSize 20 net patternnet hiddenLayerSize The training data set is divided randomly into training 70 validation 15 and testing 15 set Scaled conjugate gradient backpropagatio
33. the highest peak within a defined time window A nega tive or positive acceleration determines the direction of acceleration along its measur ing axis Considering peaks in both directions the maximum of the absolute values is taken 3 MATERIALS AND METHODS Magnitude of acceleration vector Acceleration is recorded using tri axial accelerometers which divide the experienced impact into measurements in three spatial axes Since these axes are linear indepen dent they can be concatenated to a three dimensional vector The length of this vector represents the efficient value of the resulting acceleration and is calculated with the Euclidian norm This absolute acceleration is mainly used for easier understanding in graphical visualizations of acceleration Classifying the dynamic nature of a signal in chapter 4 2 2 will utilize this descriptive feature as well The descriptive methods above are considered as traditional features that are used for acceleration activity recognition GjGC10 and have been applied by HANSON ET AL HGP 11 PREECE ET AL PGKHO9 BOUTEN ET AL BKV 97 and many more 3 2 2 Fourier analysis Human daily activities such as walking running or climbing stairs have periodic movement patterns This suggests transforming the signal into the frequency domain and investigating its spectral components with a Fast Fourier Transform FFT In this representation the main periods are represented by non zero values
34. the network to the output layer Consequently the classification result is generated and compared to the desired output Its difference represents the error of learning During the second phase this error is propagated backwards through the network to the input layer in order to change the connection weightings in such a way as to decrease the value of the error These phases are applied repeatedly with a set of training data until the error converges or a time limit is exceeded Kram09 BaHa00 Kraw13 3 4 Accelerometer Accelerometers are sensors that measure acceleration Acceleration is the rate of change in velocity in respect to time It is a vector and has magnitude and direction Accelerometers measure in units of g force which refers to the earth gravitational ac celeration of 9 81 m s Accelerometers can measure vibrations impacts tilt angle and motion of objects SENSO8 There are numerous types of accelerometers which differ in principle of sens ing and sensing element Capacitive accelerometers sense a change in electrical ca pacitance which changes between a static and dynamic condition of the sensor A piezoelectric accelerometer is based on the piezoelectric effect Certain crystals create an electric charge as external stress like acceleration is applied Piezoresistive accel erometers measure the change in electrical resistance of a material under mechanical stress Accelerometers based on the Hall Effect sense volta
35. top graph For a better perception the re sulting absolute acceleration from all three axes is drawn and not every axis individual ly The classification results are shown in the middle A nominal scale lists the different kinds of activities Every blue dot stands for a classification In the bottom graph one can see the certainty of every classification The ANN also gives information about how sure it is about a classification Since an ANN can only classify activities which were trained beforehand signals from unknown activities are stated as unclassifiable This is the case if the classification certainty is below 80 Additionally the user can zoom in and out in each window screen All time axes are linked together i e Zooming in one window simultaneously changes the other win dows An external function datetickzoom m enables a proper zoom behavior S http www mathworks com matlabcentral fileexchange 15029 datetickzoom automatically update dateticks content datetickzoom m 47 5 IMPLEMENTATION Absolute acceleration of Tibia starting on 2013 08 26 10 51 dr T T wll il I Ml 1 1 l 52 19 53 45 55 12 acceleration g N time MM SS Activity classification of Tibia starting on 2013 08 26 10 51 jogging J stairs down j nee 2 stairs up ne walking slow hon nennen 4 2 walking normal HEED Poo er Seton eens 4 3 Iying eo o 4 lt sitting 4 standin
36. training course on page xx was completed the pink green colored template should approximately fit to the acceler ometry data static activities are coded green and dynamic activities are surrounded by pink lines the vertical lines can be dragged in the right position so the activity is framed the activity identification with the template can be saved with Save tem plate If a custom training course was completed the activity template can be disabled activities can be framed with the feature Select activity data the square framing must include the red dots which are indicating the dynamic nature of the accelerometry data blue lines of acceleration do not need to be framed press Yes if selection is done selections cannot be adjusted afterwards wrong selections have to be reframed by pressing Redo Selection Undo of selection is not possible and requires a start over from step 3 new activities can be added with ADD NEW ACTIVITY and must be named and assigned to static or dynamic activities acceleration data must not be selected twice 8 When finished press DONE 9 Do steps 6 8 with hip data 10 The new neural network is saved as a MAT file in the Training folder which was chosen in step 4 xvi 5 3 APPENDIX How to classify activities Follow the instructions to apply the classification procedure on accelerometry data 1 2 o Press the button Classification With the Browse buttons choose e a MAT f
37. umop dn MO S eunou BuIA 3 3 Issejoun Sigg saes Bumem Bunem UIA uns ulpuels x11yeWw UOISNJUOD ynss4 UO edIJISSe D 63 6 EVALUATION eq WO XLJEW UOISNJUOD OY G 9 Ae Adeinsoy jeI3AO TS 88 Aqiqeyay esesaay 5 y6 Adeinsoy asesaay VL Es Ayqey 000 000 00 0 jqe ssejpun 0 0 0 4 4 4 4 4 000 9w0 00 0 00 00 0 3u1330f 0 8 0 0 0 000 00 0 000 00 00 0 geg 0 0 0 0 0 000 000 000 000 000 dn s1te3s OT 0 0 0 0 0 99 T 6r e oro 00 0 00 T50 ssejo MoIs 3UNEM 67 19 L 0 0 6 enypy ZLT 000 6T9 00 0 00 0 000 jew Oe 0 EE9 0 0 0 40u ZumjeM LTO 000 000 90 0 DE 0 0 0 T E90 000 00 0 00 0 E9 r 980 spe TE 0 0 o 18 ST L60 00 0 00 0 000 00 0 00 0 000 Z60 909 Suen LT o 0 0 o a 90T R aey 31830 umop dn Mo s guou Issejoun ege s eIs saes ZumjeMm Sunem 000 00 0 0 0 S ly o Ave na9y BulAq Zus Suipueys xew UOISN UOI 3 nsaJ uolyesiyisse 64 7 DISCUSSION AND CONCLUSION 7 Discussion and conclusion 7 1 Discussion Discussion of evaluation results In order to evaluate the developed classifier confusion matrices of classification results were created These matrices give some indication of misclassifications and perfor mance parameters Most activity activity misclassifications
38. was valued as time efficient Req8 Furthermore it is possible to work with the software program on any computer system without MATLAB license as a standalone executable Req6 In addition the software program is usable beyond the NutriHEP study and allows activity analysis with a single accelerometer from Gulf Coast Data Concepts Req9 and Req10 Accuracies of 87 99 and 71 23 were approached in classifying the activity protocol and the free run respectively Conse quently the required accuracy of more than 80 is only partly approached but still considered as entirely sufficient SysReq5 7 3 Outlook Further experimental investigations will be needed to estimate the suitability of the se lected features in this work A principal component analysis may identify the most prominent features and allows reducing the number of features Fewer features can reduce the computational complexity and remove possible redundancy in features without impairing the classification results Additionally other signal describing meth ods may be added to the pool of features to improve the performance of the classifier The evaluation in this thesis is based on a limited choice of activities The activi ty catalog represents everyday activities which are assumed to be performed frequently during the NutriHEP study New activities e g cycling driving or working may be added to the activity catalog during the evaluation Furthermore a training and ev
39. 08 the most suitable model is of order 3 autoregressive coefficient AR estimating AR model coefficient based on Burg s method with order 3 ARX arburg winX 3 ARY arburg winY 3 ARZ arburg winZ 3 The SMA is calculated according to Equation 3 11 in chapter 3 2 3 Basic arithmetic operations are used as shown below signal magnitude area SMA SMA sum abs winX sum abs winY sum abs winZ 5 3 5 Neural network training Function name main_train m This chapter explains the main function for the training of an ANN from acceleration data It is adjusted for the use during the NutriHEP study thus combining the training of two ANNs static and dynamic for two accelerometers hip and tibia According to Req9 a training function for a single accelerometer is available It is based on the same algorithm principles and the code is adjusted to one ANN for one sensor position The main function is called main train single m A standardized activity catalog and its number of static activities are declared as global variables The user chooses a folder containing the training data from hip and tibia position It is assumed that the second upper folder name corresponds to the sub ject s ID which is ensured by using the import function see chapter 5 2 1 The ANNs for the tibia and for the hip are trained and saved 40 5 IMPLEMENTATION Initiation of global variables At first two
40. FFT is calculated and the three highest peaks are considered Their magnitude and frequency are saved in a feature matrix Calculate features for every moving window The transformation into the frequency domain is approached with the built in t func tion The first component of the result is the sum of the data and can be removed Math13a The length of time in seconds of the window is calculated As described in chapter 3 2 2 the power of a signal is the squared Fourier transform The scaling fre quency goes up to the length of time in Hertz With the help of the built in function 38 5 IMPLEMENTATION findpeaks the first three peaks in power and the corresponding frequencies are de termined They are saved in a feature vector in descending order of the power ZreallesERT f fft winData ely LA n length f calc time length of acquisition Tl nt race calc power of FFT analysis power abs f n 2 power power 1 floor n 2 2 freq 1 n 2 T find peaks pks locs findpeaks power sortstr descen pks 1 find ks y Sta a eu save first three peaks sorted in feature array Npeaks 3 feat _ vect pks 1 Npeaks freq locs 1 Npeaks feat mat feat matfeat vect Function name calcKhan_movWin m This function calculates the following features from accelerometry data SE AR coeffi cients and SMA All calculations are done according to KHAN ET AL KLLK10 There
41. RUPRECHT KARLS UNIVERSIT T HEIDELBERG HOCHSCHULE HEILBRONN Medizinische Informatik Human Activity Pattern Recognition from Accelerometry Data Master s Thesis written and presented by Dipl Ing FH Dennis Jos in fulfillment of the requirements for the academic degree of Master of Science completed at the German Aerospace Center Deutsches Zentrum f r Luft und Raumfahrt e V DLR Institute of Aerospace Medicine Department of Space Physiology Cologne Germany supervised by Prof Dr Oliver Kalthoff Hochschule Heilbronn Dr rer nat Uwe Mittag DLR Cologne November 2013 UNIVERSIT T or Deutsches Zentrum HEIDELBERG DLR f r Luft und Raumfahrt ZUKUNFT SEIT 1386 German Aerospace Center ABSTRACT Abstract English Title of master s thesis Human activity pattern recognition from accelerometry data Author First examiner Dipl Ing FH Dennis Jos Prof Dr Oliver Kalthoff Second examiner Dr rer nat Uwe Mittag Ambulant studies are dependent on the behavior and compliance of subjects in their home environment Especially during interventions on the musculoskeletal system monitoring physical activity is essential even for research on nutritional metabolic or neuromuscular issues To support an ambulant study at the German Aerospace Center DLR a pattern recognition system for human activity was developed Everyday activi ties of static standing sitting Iying and dynamic nature walking asce
42. Tri axial accelerometers are attached to the orthosis and the subject s hip to acquire data of acceleration These data sets are used to assess differ ent daily physical activities since accelerometry is preferred because acceleration is proportional to external force and hence can reflect intensity and frequency of human movement YaHs10 Classification methods are then applied to recognize patterns of human activity in the acceleration data Several solutions with different methods for this process do exist but no gold standard can be identified Therefore one aim of this the sis is to find a new combination of methods for activity recognition The other aim is to develop of a human activity recognition software program which will be ready to oper ate during the NutriHEP study In conclusion this master s thesis focuses on the definition and implementation of an analyzing method which uses accelerometry data to identify characterize and allocate acceleration patterns to human physical activity 1 2 Thesis outline In chapter 2 the user requirements are acquired and derived into system requirements Chapter 3 is dedicated to explaining the materials and methods used in this thesis and summarizing the state of research Chapter 4 deals with decisions towards the archi tectural design and the data flow is visualized and explained in detail In chapter 5 the implementation of the concept into an operable software program is describe
43. a sensor starting on 2013 08 15 14 52 00 00 time MM SS BEE worn MA not worn Figure 5 4 Compliance results GUI This graphical visualization of the results is mainly intended for looking at a qualitative compliance profile It is plotted with the function plotComp1 m 5 6 Standalone executable The MATLAB Compiler assists in transforming programs into standalone applications Created applications use the MATLAB Compiler Runtime MCR which allows royalty free deployment on computers without any valid MATLAB license An installer for the MCR is packaged with the application The standalone executable does not need any installation and can be started after copying The GUI of the standalone application looks like the original software program except for the buttons for training Unfortunately MATLAB Compiler is only able to compile pre trained network command line functions Math13d l e training a new ANN with a standalone application is not possible Therefore the training functions are automatically disabled 51 6 EVALUATION 6 Evaluation During the program development algorithms were already tested with data acquired by a single person In order to evaluate the practical suitability of the program data of several subjects were required With the help of this evaluation the applicability of the program was rated for its practical use The results of the evaluation may be used for further changes to the prog
44. accelerometer The second accelerometer was worn in a belt bag see Figure 6 2 Thereby the sen sor was placed at the hip and closer to the center of mass of the body It was attached with Velcro tape so the sensor could not roll along its longer axis It was checked that the sensor kept its orientation in the belt bag throughout the whole evaluation The ori entation was consistent for all subjects On the tibia position an X6 2 accelerometer and on the hip position a X6 1B ac celerometer were attached The accelerometers were set to a sample rate of 20 sam ples per second a low gain and a 12 bit resolution How to configure the accelerome ters and how to use the analyzing software is documented in the manuals of the accel erometers GCDC10a GCDC10b and the user manual of the software see appendix 9 2 54 6 EVALUATION For video documentation a GoPro Hero 3 Silver Edition camera was mounted to a chest mount harness see Figure 6 3 b The camera was only used during the third run where the subject chose an individual course Its view was sufficient to man ually identify the types of locomotion of the subjects see Figure 6 3 c a c Figure 6 3 Subjects wearing tibia and hip sensor a GoPro chest mount harness b and GoPro camera point of view c 55 6 EVALUATION 6 2 2 Activity classes Subjects had to perform monitored movements in the laboratory and in the building respectively More precisely
45. addtodate startTime floor rawT i 1000 millisecond end Import and convert rest of accelerometer raw data The remaining CSV files are imported and converted in the same manner as before Start time information is retrieved from every single header An abortion of the conver sion process is possible by closing the loading progress bar window Plotting graphs of acceleration If a graphical visualization is requested absolute acceleration in one window and all axes individually in the other window are plotted To this end the purpose built func tions plotAbs mand plotXYZ m were programmed see appended CD 36 5 IMPLEMENTATION 5 3 2 Dynamic nature classification Function name dyn_natureClassification m This function classifies the dynamic nature of an acceleration signal whether it is of static or dynamic nature To this end the SD of the absolute acceleration is calculated If a certain threshold is exceeded the activity is considered as dynamic Classification using standard deviation The threshold for differentiating between dynamic and static activities is set to 0 1 g For every moving window the SD of the absolute acceleration is calculated with the built in std function With an if else statement the dynamic nature is saved as static 0 or dynamic 1 Additionally the time stamp in or next to the middle of the data window is saved threshold 0 1 calc SD of absolute acceleration if std winData lt
46. alua tion set of acceleration data in a more practical environment may have enhancing ef fects A subject could be sent home for one or two days wearing a camera mounted to the chest The activities recorded by the camera are no longer performed under labora tory conditions This may lead to better reliability and accuracy parameters A further issue to resolve for future work is the classification of static activities The demonstrated classifier has problems with distinguishing between sitting and standing This is acceptable for the use during the NutriHEP study because sitting and standing have similar effects on the leg muscles However future use may require a distinct differentiation Further work on feature selection and accelerometer positioning would help resolving this problem Concerning the accelerometer position sensor displacement is an important issue for future research Since there is a chance that the sensor orientation may shift and attachments may loosen enhanced sensor attachments are needed or a method for orientation independent acquisition must be developed 68 7 DISCUSSION AND CONCLUSION For a wider use of the software program future work should enhance the ad justability of the classifier Most algorithm variables are hardcoded and can only be changed by a programmer Instead a few variables should be adjustable by the user via the GUI These variables may involve the number of hidden neurons window sizes ove
47. ampling rates high frequency noises are present and need to be filtered out When a lower sampling rate is chosen no antialiasing filter is necessary on the accelerometry data PGKHO9 A moving average filter of order three was incorporated by KHAN ET AL KLLK10 to filter out random noise But with a sampling rate of 20 Hz an average filter would smooth signal peaks which may contain important information about the acceleration within an activity Therefore no moving average filter is applied In conclusion the pre processing contains the conversion of the accelerome ter s raw count data into g force data according to the GCDC user manual GCDC 10a No kind of filtering is required due to a low sampling rate of 20 Hz which is sufficient to asses daily physical activity BKV 97 4 2 2 Dynamic nature classification Windowing The signal is divided by 50 overlapping moving windows with a size of 6 seconds In this work human activity like walking or climbing stairs is considered as physically effective within a minimal lasting time of 6 seconds Furthermore a larger window size results in a less precise time allocation of activities However a small window size is unrewarding because low numbers of samples affect the frequency analysis An over lap of 50 was proven to be successful by RAVI ET AL RDMLO5 and LEE ET AL LKK 10 Additionally unpublished tests with the window size were established and came to the same result dur
48. an theory An artificial neural network ANN has a simi lar structure It is a network of processing units with weighted connections to each oth er Kram09 Learning in an ANN means to adjust the weightings of the connections A dis tinction is made between supervised and unsupervised learning During supervised learning an ANN receives an example input and additionally information about the out put result Backpropagation is a supervised learning algorithm and will be explained later Unsupervised learning is based on a self organizing network without external feedback The network is able to adjust its structure according to the input data Kram09 15 3 MATERIALS AND METHODS Many types of ANNs are suited for classification problems An ANN as a classi fier requires a learning phase It needs to be trained and learn how to classify chosen features With supervised learning a training set of input data is available containing example input features with known output class In this phase the classifier learns how to classify input data correctly with the aim to be able to apply this knowledge to new input data with unknown output This ability to generalize knowledge requires a valida tion before practical use of the classifier To this end classification result and true class affiliation are compared and its error estimated It is important that this testing input is not a subset of the training input Too much training may result in
49. ave been studied insufficiently there are no satisfying therapies or precautions available yet Consequently counter actions are required to restore health and physical capa bility of astronauts during long duration space missions Research in this field is pro gressing with immobilizing studies on earth where the loading of muscles and bones is reduced for a period of time This unloading intervention can be applied either with bed rest to the whole body or partly with a custom lower leg orthosis Conditions and meta bolisms in muscles and bones are monitored before during and after the intervention Whereas during bed rest studies the subject is not allowed to get out of bed an ortho sis study leaves the opportunity for the subject to move freely because an ambulant setting is feasible Latter aspect involves that the daily physical activity of the subject is unknown or can only be protocoled by the subject herself himself Since physical acti vity has an influence on muscle and bone condition there is a need to monitor the sub ject s daily physical activity This master s thesis is conducted on behalf of the German Aerospace Center DLR in Cologne Germany In 2014 the Department of Space Physiology at the Insti tute of Aerospace Medicine will start the NutriHEP study This study investigates the influence of nutrition and neuromuscular stimulation on the local insulin sensitivity in the calf muscle during immobilization Subjects ar
50. cause same inputs always give the same results as long as the weightings do not change Consequently the randomness concerns the training process of the ANN ANNs are initialized with random weightings i e the same training set can lead to different ANNs However if the validation error during training converges and barely changes the effect of randomness is marginal and has no effect on the performances of an ANN ANNs have a very robust perfor mance However the process of training is hard to comprehend DeSS1 1 4 2 Data flow process The classification process with ANNs can be divided into two phases see Figure 4 1 In the first phase of training and testing acceleration data is analyzed and its features are extracted The data with known output is randomly separated into training set and testing set The weightings of the ANN are adjusted until the validation error converges The result is a trained ANN which is capable of classification of activities The second phase begins with the acquisition of new acceleration data with unknown classes Its features are extracted with the same functions as before The trained ANN is then fed with these input features and classified activities are the result 24 e Training amp Testing Accelerometry data with known output Signal analysis Feature extraction Training of ANN Testing of ANN 4 DESIGN CONCEPT Repeat until validation error stops decreasing Tra
51. d and the major functions and their source codes are covered The evaluation of the software program follows in chapter 6 Its results and the course of this thesis are discussed in a final conclusion in chapter 7 2 REQUIREMENTS 2 Requirements 2 1 User requirements The aim of this thesis is to develop a software program which will be used during the NutriHEP study The conducting researcher of this study will be also the end user of the software Therefore this person was interviewed to identify what her requirements Req and expectations were The following questions and statements of the research er imply the main aspects which were mentioned Reg When did the subject do which body movement The analysis software should have the ability to tell when and how the subject moved wearing the HEP orthosis Figure 2 1 shows an example of a daily overview It was created in collaboration with the conducting researcher The diagram shows activity as a function of time Figure 2 1 is an example of a possible outcome of the software Activity profile Day 3 Subject A Activity Running Walking Standing w orthosis Sitting Lying Others gt Sd D D di PHP HP HP HP HP HM HP d HF di HF J SF VE SS SG DP E A BS GO ay O Figure 2 1 Example of graphical visualization for Req1 2 REQUIREMENTS Req2 How long did the subject do which body movement The analysis software should have the ability to tell
52. e approached by LEE ET AL LKK 10 Their system consisted of three ANNs state recognition static activity recognition and dynamic activity recognition In this thesis a similar structure is used Instead of the state recognition ANN a SD decision rule is implemented This lowers the complexity of the classifier and the training process The settings of the ANNs basically go according to HANSON ET AL HGP 11 The hidden layer has a size of 20 neurons HANSON ET AL stated that a difference of 5 neurons did not affect the results The training data set is divided randomly into training 70 validation 15 and testing 15 set Scaled conjugate gradient backpropagation is used as training function and the mean square error is the perfor mance parameter which has to converge 30 4 DESIGN CONCEPT Compliance check A decision rule is incorporated to estimate the subject s compliance The check for compliance is only applicable during the NutriHEP study In collaboration with the con ducting researcher scenarios were elaborated concerning the wearing of the HEP or thosis With the activity classification described above a differentiation between static and dynamic activities is possible Looking at one time window there are four possible combinations of classifications 1 Tibia and hip sensor both measure dynamic activities 2 Tibia and hip sensor both measure static activities 3 Tibia sensor measures dynamic activity and
53. e partly immobilized by wearing the custom built HEPHAISTOS HEP orthosis on one leg This orthosis developed by WEBER ET AL WDM 13 unloads the muscles of the lower limb while retaining the ap plication of body weight on the skeletal structure It is designed for ambulant studies and allows everyday locomotion This directed immobilization of the calf muscle in duces muscle atrophy which affects insulin sensitivity In the NutriHEP study electrical stimuli and a specific diet with lupine seeds will be tested for their effectiveness on in sulin sensitivity Any influence on insulin sensitivity may affect the degree of muscle loss It is required to validate the relationship between muscle loss and orthosis inter vention A possible side effect of wearing the orthosis may be that the subject is less active This whole body inactivity can call systemic effects on the insulin sensitivity of the muscles WSW 13 To locate interventional effects on the calf muscle only it is essential to show that the physical activity of the whole body remains the same Activity logging by the subject is neither reliable nor accurate Therefore an ambulant monitor 1 INTRODUCTION ing system of the physical activity is needed Furthermore a documentation of the sub ject s compliance is required to ensure that the orthosis was worn throughout the study For an ambulant study accelerometers are a low cost and practical solution to monitor physical activity
54. e you through several functions and their purposes for the NutriHEP study 5 1 How to synchronize two accelerometers For a proper classification with tibia and hip sensor the two accelerometers need to be synchronized since the accuracy of the time setting as in 3 2 Mandatory settings of accelerometers for the use with MovelT on page xii is insufficient To synchronize two accelerometers a shared impact on both sensors is needed 1 Turn both sensors ON Slightly bump sensors against each other Turn both sensors OFF Connect first sensor to PC via USB Open USB drive and go to GCDC folder Copy last edited CSV file Open subject folder and create a new folder Time Synchronization Paste CSV file and rename it after the sensor ID e g Acc0004 9 Delete CSV file on accelerometer 10 Do steps 4 9 with second sensor 11 Open MovelT software See 2 Installation on page ix 12 Press the button Time Synchronization 13 Browse to subject folder and choose CSV files consecutively 14 If necessary the name of the sensor can be changed 15 Both accelerometry data sets from the bumping impact in step 2 will be plotted in a new window 16 Click the peak on the graph of the blue line so a little black square appears on that point and press Done 17 Mark the corresponding peak on the green line and press Done 18 The program will synchronize the two graphs and will show the result in a new window 19 If the result is unsatisfy
55. er 2009 M Krawczak Multilayer Neural Networks A Generalized Net Perspec tive Cham Springer 2013 J R Kwapisz G M Weiss and S A Moore Activity recognition using cell phone accelerometers ACM SIGKDD Explorations Newsletter vol 12 no 2 pp 74 82 2011 M W Lee A M Khan J H Kim Y S Cho and T S Kim A single tri axial accelerometer based real time personal life log system capable of activity classification and exercise information generation in Engineer ing in Medicine and Biology Society 2010 EMBS 2010 32nd Annual International Conference of the IEEE pp 1390 1393 2010 XIV LoYA09 Math13 Math13a Math13b Math13c Math13d MCLC04 Mull13 REFERENCES X Long B Yin and R M Aarts Single accelerometer based daily physical activity classification in Engineering in Medicine and Biology Society 2009 EMBS 2009 31st Annual International Conference of the IEEE pp 6107 6110 2009 The MathWorks Inc Fast Fourier transform MATLAB fft Mathworks Online Available http www mathworks com help matlab ref fft html Accessed 05 Oct 2013 The MathWorks Inc MATLAB Using FFT Demo Mathworks Online Available http www mathworks com help matlab examples using fft html Accessed 06 Oct 2013 The MathWorks Inc Power spectral density MATLAB dspdata psd Mathworks Online Available http www mathwor
56. eration on the screen the user has to click on the peak of the shared event for identification This has to be done for both sensors in a specific order The software will indicate that the first signal is always the blue line The times of the chosen data points are saved for the following step Calculate time shift The time shift is equal to the time difference between the saved time stamps in the step before The difference is calculated by the built in etime function which estimates the elapsed seconds between two date vectors The time difference is multiplied by a thou sand and rounded to an integer number to provide a time shift in milliseconds This conversion is needed for proper function of the following step delta_t floor etime datevec TimeA datevec TimeB 1000 Apply time shift on time array Once the time shift is known the time array can be manipulated The shifting millisec onds are added to every date vector with the built in function addtodate The smallest time dimension is millisecond and the added time has to be an integer number Visual validation and saving results For a visual validation by the user both time signals are plotted again with one being shifted If the results are unsatisfying i e the peaks of the shared event are not match ing in time the time synchronizing process has to start over If the results are satisfy ing the calculated time shift is saved in a DAT file for future processin
57. etry sensors from Gulf Coast Data Concepts xi APPENDIX 3 1 How to use the XLR8R software The onboard configuration software XLR8R by GCDC is stored on the flash drive of every accelerometer If you cannot find the software on the flash drive download the latest version on http www gcdataconcepts com analysis html To run the standalone software JAVA 6 0 or later needs to be installed on your computer Ask your departmental IT administrator for further help Starting the XLR8R software 1 Plug in your accelerometer via USB 2 Go to the folder xIr8r and double click on XLR8R jar 3 The XLA8R software will start showing the home menu Please note For further instructions how to use the XLR8R software re fer to its help documentation stored on the accelerometer s flash drive or online on http www gcdataconcepts com XLR8R_r2 help paf 3 2 Mandatory settings of accelerometers for the use with MovelT The following settings are mandatory for MovelT working properly Setting time 1 Under Utilities open the tab Set Device Time Press the button Write File Immediately close the XLR8R window and safely unplug the accelerometer Start the accelerometer right away After a few seconds you can stop the acquisition Why do I have to be so quick When the button Write File is pressed the current computer time is saved in a txt file on the accelerometer On the next boot of the accelerometer the time in the txt f
58. g urn A ech unclassifiable i 55 00 time MM SS Classification certainty of Tibia starting on 2013 08 26 10 51 z 100 GE D 50 D 0 55 00 time MM SS Figure 5 2 Classification result GUI This graphical visualization of the results is mainly intended for looking at specific sin gle classifications and the raw acceleration signal It is plotted with the function plotClassification m 48 5 IMPLEMENTATION Pie chart A visualization of the relative distribution of activities is given by a pie chart see Figure 5 3 Its percentages are relative to the total time of classification The occurrences of activity are counted and plotted with the built in pie function Activity distribution of Tibia 2 2 HB unclassifiable HEB standing gt Dying Fa walking normal MG stairs up MA stairs down HEB jogging Figure 5 3 Pie chart results GUI 9 i This graphical visualization of the results is mainly intended for looking at a qualitative activity distribution profile It is plotted with the function plotPie m 49 5 IMPLEMENTATION Table A table with numeric information is generated see Table 5 1 At the top it shows the elapsed time of the classified acceleration data Beneath the average certainties of dynamic and static classifications are written This information helps to estimate the quality of the classification In the lower half all ac
59. g It will be used during the activity classification in chapter 5 4 1 34 5 IMPLEMENTATION 5 3 Training phase This chapter explains the implemented code in MATLAB for the software functions which are necessary during the training phase of the ANNs Only fundamental code lines are presented here The full program code is accessible on the appended CD 5 3 1 Data conversion Function name multiple_RAW_Conversion m This function executes an import of all CSV files in one folder The raw data is convert ed into g forces If more than one file is present the alphabetical order must corre spond to the chronological order of the files First the number of input arguments is checked since this function has default inputs If no pathname is given a folder has to be selected The first file is imported and its header information is retrieved The con version from raw count data into g force data follows the instructions in the manual from Gulf Coast Data Concepts Finally the absolute acceleration is calculated and if needed everything gets plotted for visual representation Input validation Two input variables are configurable plot request and pathname If a plot is requested this variable has the value 1 default value Any other value will not result in a plotting screen at the end of this process The pathname defines the folder containing the data files for conversion If no pathname is given it has to be selected via a pop up
60. g questions as detailed as possible Feel free to just write single words notes or drawings Program software 1 What is your first impression of the software Easy to understand Time efficient analysis of data Clear output 2 What further functions or features do you miss Nothing yet Diagrams 1 Which diagram do you prefer and why All diagrams serve a purpose There are diagrams for just getting a good overview and diagrams to get information in more detail 2 Is there any information which you see as unnecessary No 3 Which information or graphics do you miss None Page 1 of 4 APPENDIX User manual 1 What do you like about the user manual Most processes are described very clearly and are understandable in combination with using the software Figures help to understand the processes where necessary No unnecessary information 2 What do you dislike about the user manual Some points are missing gt therefore not always super clear how to proceed Any further comments Missing points in the manual were discussed with the author and will be resolved by the author Page 2 of 4 APPENDIX 1 2 Questionnaire Please answer the following questions by ticking the appropriate scaling Program software 1 The menu looks neat totally partially partially totally disagree disagree agree agree O O L
61. ge variations in the mag netic field around the sensor Magnetoresistive accelerometers have a similar structure and function but instead of measuring voltage they measure changes in resistance due to a magnetic field Accelerometers based on micro electromechanical systems MEMS technology are small structures with dimensions in micrometer scale This technology is now being utilized to manufacture state of the art accelerometers SENS08 GCDC accelerometer As stated in Req10 the type of accelerometer to use is determined Gulf Coast Data Concepts GCDC sell accelerometers based on MEMS The 3 axis digital accelerome ter sensors are manufactured by Analog Devices Norwood MA USA They are typi 17 3 MATERIALS AND METHODS cally accurate to within 10 using the conversion methods provided in the user manu als GCDC10a GCDC13 Most of these errors are due to an offset error e g when the sensor in rest reports gravitational acceleration of 1 1 g instead of 1 0 g Figure 3 2 USB Accelerometer X6 2 GCDC Waveland USA During software development an USB accelerometer model X6 1A is used for the hip position and an USB accelerometer model X6 2 see Figure 3 2 is used for the tibia position Both have the same dimensions but different types of power supply The X6 1A is powered by a single AA sized alkaline battery while the X6 2 has an inter nal hardwired 500 mAh lithium polymer rechargeable battery GCDC10b Si
62. he 23rd International Technical Meeting of the Satellite Division of the Institute of Navigation pp 2906 2912 2010 M Frigo and S G Johnson FFTW An adaptive software architecture for the FFT in Proceedings of the 1998 IEEE International Conference on Acoustics Speech and Signal Processing vol 3 pp 1381 1384 1998 Gulf Coast Data Concepts LLC GCDC X6 1A User Manual Gulf Coast Data Concepts Waveland MS 2010 Gulf Coast Data Concepts LLC GCDC X6 2 User Manual Coast Data Concepts Waveland MS 2010 A Kooney Gulf Coast Data Concepts Re X6 2 Accelerometer Physi cal principle Email 05 Jul 2013 K O Genc V E Mandes and P R Cavanagh Gravity replacement during running in simulated microgravity Aviation Space and Environ mental Medicine vol 77 no 11 pp 1117 1124 2006 H Gjoreski M Gams and I Chorbev 3 axial accelerometers activity recognition ICT Innovations pp 51 58 2010 A Hein and T Kirste Towards Recognizing Abstract Activities An Un supervised Approach in Proceedings of the 2nd Workshop on Behav our Monitoring and Interpretation BMI 2008 Kaiserslautern vol 8 pp 102 114 2008 A M Hanson K M Gilkey G P Perusek D A Thorndike G A Kutnick C M Grodsinsky A J Rice and P R Cavanagh Miniaturized Sensors to Monitor Simulated Lunar Locomotion Aviation Space and Environmental Medicine vol 82 n
63. hesis the conducting research er of the NutriHEP study decided to integrate skin conductance sensors in the HEP orthosis These sensors can detect skin contact and thus if the orthosis is worn or not This technology promises a high accuracy Therefore the requirement for the compli ance check was downgraded and seen as supporting feature Discussion of usability test results In the course of this thesis only a short usability test was carried out with the conduct ing researcher of the NutriHEP study This test involved studying the user manual ful filling all possible tasks and completing a subsequent questionnaire This evaluation of usability by only one person was considered acceptable because the focus of this the sis was not on the development of a marketable software package In the usability test the program software and its use were experienced com prehensible and efficient The tester commented that at some points the program did not explain itself so that necessarily the user manual had to be consulted This point of criticism may be justified on grounds of user friendliness but since a user interface overloaded with information leads to confusion the information missing from the inter face were deliberately placed in the user manual The user manual itself was considered helpful and very clearly written Espe cially figures of screen shots were appreciated Some step by step instructions skipped certain steps and led to co
64. hip sensor measures static activity 4 Tibia sensor measures static activity and hip sensor measures dynamic activity In the first two cases the classifications seem consistent and the subject is either mov ing or not in motion at all In the third scenario it can be assumed that the hip sensor is not attached Other explanations could be a sitting subject moving the leg or doing in door cycling Nevertheless it appears plausible that in the third scenario the subject is still wearing the orthosis The fourth scenario can be interpreted as a moving upper body and static lower body This combination of body movement is difficult to realize Attempts in generating such scenarios e g with hula hoops failed because the tibia sensor always detected dynamic activities as well This finding leaves the explanation of an unattached tibia sensor In fact the tibia sensor is permanently connected to the orthosis so the fourth scenario indicates that the orthosis is not worn The decision rule is not applied to every single pair of tibia and hip classifica tion A window of 240 samples 12 seconds is checked if the fourth scenario ap plies If it applies for all pairs of samples the orthosis is considered as not worn 31 5 IMPLEMENTATION 5 Implementation This chapter describes the programing structures which were developed to implement the classification concept All tools and programs used during this thesis are listed The main feature
65. hods for the Classification of Dynamic Activities From Accelerometer Data EEE Transactions on Biomedical Engineering vol 56 no 3 pp 871 879 2009 H Pruscha Clusteranalyse in Statistisches Methodenbuch Verfah ren Fallstudien Programmcodes Berlin Heidelberg Springer pp 289 316 2006 N Ravi N Dandekar P Mysore and M L Littman Activity recognition from accelerometer data in Proceedings of the AAAI Conference on Artificial Intelligence vol 20 p 1541 2005 C Sammut and G I Webb Encyclopedia of Machine Learning New York Springer 2011 SENSR Practical Guide To Accelerometers Practical Guide To Accel erometers Online Available http www sensr com pdf prac tical guide to accelerometers pdf Accessed 08 Nov 2013 C E Shannon and W Weaver The Mathematical Theory of Communi cation Urbana The University of Illinois Press 1964 S W Smith Digital signal processing a practical guide for engineers and scientists San Diego CA California Technical Pub 2003 K M Ting Confusion Matrix in Encyclopedia of Machine Learning New York Springer p 209 2010 J V Tu Advantages and disadvantages of using artificial neural net works versus logistic regression for predicting medical outcomes Jour nal of Clinical Epidemiology vol 49 no 11 pp 1225 1231 1996 XVI REFERENCES Wags02 WDM 13 WDY 13 WSW 13 YaHs10 K
66. ierungen lag bei 96 49 bzw 76 77 Diese Leistungsparameter beschreiben ein funktionierendes ambulantes Monitoring System von k rperlichen Aktivit ten Schl sselw rter Bewegungserkennung Beschleunigungssensoren k nstliche neuronale Netze ambu lantes Monitoring berwachtes Lernen EIDESSTATTLICHE ERKL RUNG Eidesstattliche Erkl rung Universit t Heidelberg Hochschule Heilbronn Studiengang Medizinische Informatik Autor Matrikelnummer Dipl Ing FH Dennis Jos 179715 Thema der Masterarbeit Human Activity Pattern Recognition from Accelerometry Data Ich erkl re hiermit an Eides Statt dass ich die vorliegende Arbeit selbstst ndig und ohne Benutzung anderer als der angegebenen Hilfsmittel angefertigt habe die aus fremden Quellen einschlie lich elektronischer Quellen direkt oder indirekt bernom menen Gedanken sind als solche kenntlich gemacht Die Arbeit wurde bisher weder im Inland noch im Ausland in gleicher oder hnlicher Form einer anderen Pr fungsbeh rde vorgelegt und ist auch noch nicht ver ffentlicht K ln den 26 November 2013 Dennis Jos ACKNOWLEDGEMENT Acknowledgement This master s thesis was carried out at the German Aerospace Center in Cologne Support was given by the Institute of Aerospace Medicine who had funded this work in all its stages gratefully acknowledge the collaboration and help provided by Dr Uwe Mittag as supervisor and Kathrin Schopen as conducting researcher
67. ile as neural network e g NeuralNetworks_SubjectID_date mat e a folder containing Hip and Tibia folder with data for classification Define where the asked sensor is placed If the error message No time synchronization file found appears follow step 20 in 5 1 How to synchronize two accelerometers on page xv and start over Pop up windows will show the classification results e Window with tibia results will appear exactly behind window with hip re sults The window Visibility of graphical representations offers to choose which pop up windows are shown on the screen To end the classification progress press Close All The graphical representations are saved as FIG files in a new folder Graphical representations mm o 8 o gt ES gt E O lt Iki P walking normali lute acceleration of the Absolute acceleration of Tibia starting on 2013 08 26 10 16 I Wii lk acceleration g time MM SS On the left of the screen the Activity classification of Tibia starting on 2013 08 26 10 16 results are represented in a RE time dependent diagram The stairs down e stairs up upper curve shows the abso siting measurement in g force The standing etese ogansesees unclassifiable graph in the middle has a nominal scale with dots showing successful classifi Classification certainty of Tibia starting on
68. ile is loaded into the accelerometer s real time clock This means that the time between pressing Write File and a completed reboot is equivalent to the absolute time shift of the accelerometer s internal clock 6 The new generated data file in the GCDC folder of the accelerometer can be deleted HUA SS Pe xii APPENDIX Subject ID and sensor position 1 Open the XLA8R software XLR8R jar 2 Under Utilities open the tab Configuration File Editor 3 In the Comments section type in subject ID and sensor position in the following manner Comments subject ID sensor position It is important that the subject ID appears in the first row and the sensor posi tion in the second row e For the NutriHEP study write as sensor position Hip or Tibia 4 Press Save File 5 Close the XLR8R window 4 Getting started with MovelT The MovelT software offers several features for gaining information about a person s activities The following documentation will guide you through the preparing proce dures 4 1 How to set the default folder path In MovelT files and folders often need to be chosen in a browser window To avoid too much clicking along the same paths you can define a default folder where the browser window should start There are two ways to change the default folder path 1 Under Default Folder Path press Browse 2 Click through to you folder of choice 1 Under Default Folder Path click in
69. individually This uneven distribution would distort the confusion matrix and is consid ered as inexpedient The evaluation with confusion matrices showed that the accuracy of the activity course classification lied above 85 and its results were in over 95 of the cases reliable see Table 6 3 This was true for hip and tibia classification of the data set based on the completion of the activity course For the classification of the free run accuracy and reliability were over 70 and 75 respectively Table 6 3 Accuracies and reliability of classifier mean SD Overall accuracy Average reliability Activity course 87 99 7 54 96 49 3 16 Free run 71 23 16 55 76 77 16 26 An approximate computation time of 6 minutes per measurement day was measured This is true for acceleration acquisitions with a sample rate of 20 Hz 62 6 EVALUATION diy wou xuyew UOISNJUOD OY 7 9 aqe L Ave naay jeI8AO Angels adesany Angeten Ave naay a8e13ny ierg Si o o Si 0 0 00 0 00 0 000 sjgeyissejpun 0 0 0 u33or 99 T 0 ma Er en 0 m 0 Se o o 0 a o UMOP S11 S dn saieqs e I Oo 000 ITO es ad o 00 0 ES ITO sse o Mojs SUIYJEM engoy ewou ZumjeM e 00 0 SC 09 SE 000 Ke ch o L tv9 0 880 1 000 000 000 L6 S des o 00 0 o 0 0 80T 0 1 o 000 00 0 0 2 y e T Zus o 0 o 990 90 0 Sisi ras H Ave naay 21929 3u1330f
70. ined ANN e Classification Accelerometry data with unknown output Signal analysis Feature extraction Gr Classification with ANN Classified activity La a a il Figure 4 1 Training and classification concept For each subject acceleration data is acquired from two sensors placed parallel to the tibia and close to the hip Both data sets run through the same activity classification algorithm which is shown in Figure 4 2 as a flow diagram The consecutive processes of data acquisition dynamic nature classification feature extraction and activity classi fication are visualized 25 4 DESIGN CONCEPT Data acquisition Acceleration data 7 N IN V i Conversion raw data gt g force Dynamic nature classification Divide data into 50 overlapping windows with 6s length Dynamic nature For every window calculate describes how far the signal fluctuates from E dynamic nature of activity J its mean Standard deviation of absolute acceleration SD_abs Static activity gt Yes SD_abs _No lt 0 1 Dynamic activity E x z 2 s For all static phases oi a For all dynamic phases ei calculate features Calculate features S y oO oO Time domain analysis Ti
71. ing start over from step 12 20 If the result is satisfying close the window 21 In the folder Time Synchronization a new TXT file is generated containing in formation about the time synchronization e This file needs to be copied to the designated folder before classification See_5 3 How to classify activities on page xvii QO TD SO E do XV APPENDIX 5 2 How to train a neural network At this point accelerometry data should be acquired from activities which are later to be classified These data files can be saved with the Import feature as described in 4 2 How to import data on page xiv Please note A MATLABO license is needed to use the training feature Without a MATLABO license it is only possible to work with pre trained neural networks Therefore the feature is disabled To accelerate the progress the subject can complete a default training course as it is described in Attachments on page xx This course contains several activities in a spe cific order To train a new neural network follow the instructions below 1 2 3 4 an 7 Import the training data using the Import feature of MovelT Change the folder name of the calendar week into Training Press the button Training on the home screen of MovelT Browse to subject folder and choose the Training folder Chosen folder must contain folders named Hip and Tibia The opening window is for identifying activities in tibia data If the 7 1 Default
72. ing as jogging 30s u Stop at end of corridor turn around Interstation activity 90s and wait for 11 30 j j Walk slowly across corridor 6 11 30 Walking with slow pace 30s T Stop at end of corridor go to next Interstation activity l l 60s station and wait for 13 00 Ms Sit down on chair with both feet 7 13 00 Sitting 30s touching ground SR Get up from chair go to next station Interstation activity 30s and wait for 14 00 Lie down on couch with face up and 8 14 00 Lying 30s without crossing legs vk Get up from couch turn accelerome Interstation activity 30s ters off 58 6 EVALUATION 6 3 Test data processing 6 3 1 Training and classification With the acceleration data from the first run the ANNs were trained and the resulting network data saved How to use the training feature of the software is explained in its user manual See appendix 9 2 For each subject an individual ANN was trained The template feature helped to save time while training 14 different ANNs The data from the second and third run were classified by the analyzing feature Its results were then sighted for evaluation 6 3 2 Evaluation results Randomization The first and the second run were identical in their sequence and therefore inter changeable One served as training set the other one was for classification However doing the activity course twice might have had an effect on the locomotion It was ob served that
73. ing the work with WEBER ET AL WDY 13 The following calculations are applied to every window data Classification of dynamic nature of activity Dynamic nature describes the dispersion of an acceleration signal The dynamic nature of a human activity can be either static or dynamic Static activities are activities where no movements of the interested body parts are noticeable e g sitting or standing Dynamic activities are activities where cyclic movements can be assumed e g walk ing or climbing stairs This differentiation is needed because feature extraction meth ods in the frequency domain are not applicable to acceleration signals of static nature The SD of the absolute acceleration is considered as an easy to calculate parameter The absolute acceleration is the effective resulting value of all three spatial axes and is approached by using the Euclidian norm calculation see chapter 3 2 1 SD was prov en to be a good quantifier of an activity s dynamic nature by FRANK ET AL FNRA10 and WEBER ET AL WDY 13 because the SD correlates with the dispersion of a signal 28 4 DESIGN CONCEPT A threshold is needed to distinguish the two types of dynamic nature During the work of WEBER ET AL task related mean SDs for several activities were identified Un published observations have shown that an SD threshold of 0 1 g is appropriate to dif ferentiate between static lt 0 1 g and dynamic gt 0 1 g activities This decision
74. ity course which will be explained in chapter 6 2 3 The user can adjust these templates to the acceleration signal With one click all these templates are saved If necessary additional activities can be added and identified for the training of an ANN Plot graph with absolute acceleration and dynamic nature classification After importing the acceleration data the absolute acceleration from all three axes is plotted as a blue graph Within this graph the dynamic nature is also visualized Red dots distinguish between static 0 and dynamic 1 activities see Figure 5 1 The pink and green vertical lines represent the activity template They can be dragged by the user to allocate the flagged activity Flags are at the top of every second vertical line indicating the beginning of an activity Pink lines are dynamic and green lines are static activities When the template is saved a new folder is created named TrainingData It contains sub folders called static and dynamic In each folder the associated activities are saved Acceleration data of one activity is saved together in one file which is given the activity name 42 5 IMPLEMENTATION Absolute acceleration from tibia starting on 2013 08 14 10 12 3 5 T 3 E 2 57 4 a c 2 PRER 2 k EI 8 15 1 tz Wie 4 absolute acceleration dynamic nature 25 40 Figure 5 1 Screen for allocating during training phase with activity
75. ks com help signal ref dspdata psd html Accessed 07 Oct 2013 The MathWorks Inc Autoregressive AR all pole model parameters estimated using Burg method MATLAB arburg MathWorks Deutsch land Mathworks Online Available http www mathworks de de help signal ref arburg html Accessed 05 Nov 2013 The MathWorks Inc MATLAB Compiler 5 0 Support for MATLAB and Toolboxes MathWorks Deutschland MATLAB Compiler 2013 Online Available http www mathworks de products com piler supported compiler support html Accessed 20 Nov 2013 M J Mathie A C Coster N H Lovell and B G Celler Accelerome try providing an integrated practical method for long term ambulatory monitoring of human movement Physiological Measurement vol 25 no 2 p R1 2004 D M ller Wichards Diskrete und schnelle Fourier Transformation in Transformationen und Signale Wiesbaden Springer pp 161 199 2013 XV Neub12 OpSc10 PGKHO9 Prus06 RDMLO5 SaWe11 SENS08 ShWe64 Smit03 Ting10 Tu96 REFERENCES A Neubauer Schnelle Fourier Transformation in DFT Diskrete Fou rier Transformation Vieweg Teubner Verlag pp 125 139 2012 A V Oppenheim and R W Schafer Discrete time signal processing Upper Saddle River Pearson 2010 S J Preece J Y Goulermas L P J Kenney and D Howard A Com parison of Feature Extraction Met
76. ld be straightforward and intuitive In addition the re searcher proposed that there might be a lack of time during the data analysis of the study Therefore a colleague with no experience in data analysis should be able to work with the software 2 REQUIREMENTS Req6 The software should work as a standalone executable The preliminary requirement implies that several persons will be working with the a nalysis software Hence the software should run on different work stations A standalone executable of the analysis software ensures parallel work on independent computers Req7 The software should be provided in English and German The main language of the software should be English However it is possible that po tential users as stated in Req5 do not speak any English Thus the user should be offered to choose between English and German version of the software program Req8 The software should not be time consuming The analysis software should have as many automatic procedures as possible It should not be too time consuming because it is a supporting process of the NutriHEP study Req9 The software should work with a single sensor on any position The analysis software should be usable beyond the NutriHEP study Whereas Nutri HEP supplies the subject with two accelerometers i e tibia and hip the software should also work with a single sensor and at any chosen body position Req10 The soft
77. led frequency see Equation 3 7 shows the periodogram Math1 3a 2 Equation 3 6 adapted from Math13a Powys Yoo Y D Y E 1 0 1 1 Equation 3 7 maz N adapted from Math13a Peak magnitude Peak magnitudes are local maxima in the periodogram Their frequencies represent the most dominant frequencies in the input signal The strongest frequency Frqmain can be picked out by allocating the frequency with the highest power in the periodogram Math13a Pow Frqmain max Pow f Equation 3 8 adapted from Math1 3a Power spectral density The power spectral density PSD describes the average distribution of the signal s power in the frequency domain The average power over a frequency band is com puted by the integral of the PSD and not its peaks Math13b For time discrete sig nals the PSD is calculated from the DFT of the autocorrelation sequence Beuc1 1 It is used for the calculation of the spectral entropy SE Spectral entropy According to information theory entropy describes the sum of all microstates in a bal anced system A low number of different microstates appear in a consistent and uni form system with low entropy In contrast numerous variations of microstates cause uncertainty and confusion hence high entropy Entropy is defined as a measurement of uncertainty or complexness of a system Alve07 ShWe64 For the differentiation between deterministic and random parts of a signal en trop
78. logy laboratory 7 sitting 8 lying Figure 6 4 Map of DLR campus with building 24 and locations of the stations Risk benefit analysis and precautions Since there is no medical intervention the only risk lies in the completion of the activity course Its demanding activities are elemental types of locomotion of everyday life thus not containing any serious risks Therefore no medical advisor or paramedic is needed during the evaluation 57 6 EVALUATION Table 6 2 Order of activity course Starting FR A No Activity Description Duration time Turn accelerometers on stand in 1 0 00 Standing 30s front of stairs and wait for 0 30 eae 8 8 Walk 4 floors upstairs in a uniform 2 0 30 Climbing ascending stairs 90s pace CH Stand still at top level turn around Interstation activity 30s and wait for 2 30 ae Walk 4 floors downstairs without 3 2 30 Climbing descending stairs K 90s skipping or jumping Va Stop at end of stairs go to next sta Interstation activity 60s tion and wait for 5 00 Walk in own normal pace 4 5 00 Walking with normal pace 90s l E Stop at end of corridor turn around Interstation activity 30s and wait for 7 00 Walk back in own normal pace 4 7 00 Walking with normal pace 90s ER Stop at end of corridor go to next Interstation activity 60s station and wait for 9 30 E Jog in own slow pace 5 9 30 Runn
79. me domain analysis E Frequency domain analysis D LL D Standing Walk normal 2 o Ki Sitting Walk slow gs gt E Lying Stairs up lt Stairs down Jogging Figure 4 2 Data flow diagram 26 4 DESIGN CONCEPT 4 2 1 Data acquisition Accelerometer MEMS accelerometers are used from GCDC see chapter 3 4 Their factory calibration with 10 accuracy is sufficient because acceleration values or more precisely ex tracted features are compared relatively against each other An offset error which af fects the entire measurement does not have an effect on the classification results GCDC write in their user manual that a lower sampling rate extends the operat ing time of the sensor GCDC10b This suggests keeping the sampling rate as low as possible BOUTEN ET AL BKV 97 found frequencies up to 20 Hz to be sufficient to as sess daily physical activity using body fixed accelerometers Further studies reviewed in BKV 97 recommend an ability to register accelera tions within an amplitude range of 12 g to 12 g for ankle placed sensors and a level of 6 g to 6 g at waist Both accelerometer models from GCDC are capable only of the latter range This would indicate a problem concerning the sensor at the tibia position However these specifications are based on data acquired by BHATTACHARYA ET AL during tread mill runs with speeds up to 11 km h BMSG80 In this thesis light jogging is the only ty
80. meters attached to footwear and waist measured activities like walking and running under earth gravity Furthermore walking running hopping and loping under 1 6 earth body weight simulated activities under moon conditions Ground truth data was established by demanding certain activities Features like signal peak spread frequencies power and centroid were used as input for an ANN It achieved an accu racy of 100 KWAPISZ ET AL KwWM11 developed an activity recognition system with phone based accelerometers Via a custom built smart phone application data was acquired from the built in accelerometer It was able to recognize walking jogging climbing stairs sitting and standing For generating a ground truth data set the subjects were asked to label every activity using the installed application Mean SD root mean square time between peaks average absolute difference and binned distribution were the features of choice KWAPISZ ET AL compared decision trees logistic regression and ANNs as classifying method These methods reached accuracies of 85 1 78 1 and 91 7 respectively 22 4 DESIGN CONCEPT 4 Design concept Different body movements cause different accelerations on several body parts Every movement appears to have its individual characteristic pattern in acceleration An iden tification of acceleration patterns and their proper allocation to human activities can be approached with several classification
81. methods see chapter 3 3 This chapter ex plains the choice of the classification method and the data flow process of the activity recognition 4 1 Classification with artificial neural networks ANNS are the method of choice for human activity pattern recognition In the following it will be explained why ANNs are more suitable in this thesis than the other classifica tion methods described in chapter 3 3 A Bayes classifier is based on naive independence assumptions Those can be wrong but this does not imply that the classifier will fail Instead the classification per formance will be poor because the features tend to have a similar distribution and differences are only in relations of the features A Bayes classifier is not suitable be cause there is no knowledge about the feature distribution and the relations Their neg ative effects on the classification performance cannot be ruled out B6hm03 LONG ET AL found out that in general a decision tree has the best performance in activity classification However from a practical point of view this classification is not very suitable because it needs a lot of implementation If new activities need to be in corporated a completely new decision tree classifier needs to be re built Manual tun ing by experts might also be necessary because tree training is completed on isolated activity events This results in a low level of extensibility LoYA09 A main disadvantage of cluste
82. n Heidelberg Springer pp 399 542 2011 C V Bouten K T Koekkoek M Verduin R Kodde and J D Janssen A triaxial accelerometer and portable data processing unit for the as sessment of daily physical activity IEEE Transactions on Biomedical Engineering vol 44 no 3 pp 136 147 1997 A Bhattacharya E P McCutcheon E Shvartz and J E Greenleaf Body acceleration distribution and O2 uptake in humans during running and jumping Journal of Applied Physiology vol 49 no 5 pp 881 887 1980 C B hm Knowledge Discovery in Databases Lecture Ludwig Maximilians Universit t M nchen Munich 2003 W Y Chung A Purwar and A Sharma Frequency domain approach for activity classification using accelerometer in Engineering in Medicine and Biology Society 2008 EMBS 2008 30th Annual International Con ference of the IEEE pp 1120 1123 2008 Xl Cimi07 CoTu65 Dess11 Dobr09 Dorf13 DSD 12 DuVe90 ErPC08 FDFC10 REFERENCES P Cimiano Maschinelles Lernen Lecture Universitat Heidelberg Hei delberg 2007 J W Cooley and J W Tukey An Algorithm for the Machine Calculation of Complex Fourier Series Mathematics of Computation vol 19 no 90 pp 297 301 1965 C Demant B Streicher Abel and A Springhoff Querschnitt Klassifika tion in Industrielle Bildverarbeitung Berlin Heidelberg Springer pp 171 194 2011
83. n is used as training func tion and the mean square error is the performance parameter which has to converge Setup Division of Data for Training Validation Testing net divideFcn dividerand Divide data randomly net divideParam trainRatio 70 100 net divideParam valRatio 15 100 net divideParam testRatio 15 100 net trainFen trainscg Scaled conjugate gradient Choose a Performance Function o net performFcn mse Mean squared error The ANN is trained and tested until the mean squared error tends to a minimum Algo rithms for these procedures are provided by the Neural Network Toolbox 44 5 IMPLEMENTATION 5 4 Classification phase 5 4 1 Neural network classification Function name main m This function represents the main function for the activity classification of accelerometry data As the classification is the main feature of the software this function is called main m This function is customized for the analysis of two acceleration sensors hip and tibia as applied during the NutriHEP study The user chooses a pre trained ANN and a folder containing acceleration data The activity class catalog is saved within the ANN file because new activities might have been added If a file containing the time shift see chapter 5 2 2 is present the time shift is added to the corresponding data array First the classification for the tibia is done by calling the function activityClassification m
84. nce latter battery is rechargeable via USB cable it is used for the tibia position so the sensor can be attached permanently to the orthosis or shin guard Replacing the battery would loosen the attachment of the sensor Using a rechargeable battery ensures the sensors position and orientation The X6 1A accelerometer is able to sample data at a minimal sampling rate of 10 Hz and the X6 2 of 20 Hz Both accelerometer models are capable of measuring accelerations between 6 g and 6 g The data measured by the accel erometers is saved as raw count data According to their user manuals GCDC10a GCDC10b the conversion algorithm is dependent on the settings of the accelerome ter The gain setting changes the proportionality coefficient between count data and g force units see Table 3 1 It is common to specify acceleration not in m s but in multiples of the gravitational force g 9 81 m s Table 3 1 Converting rules from raw counts data into g units GCDC10a AD Resolution Gain Setting Deadband Counts 12 bit Low 340xg 18 3 MATERIALS AND METHODS 3 5 Evaluation methods Confusion matrix Proper function and correctness of a classifying system need to be validated and eval uated A confusion matrix is a common method to evaluate performance parameters of a multiclass classifier Ting10 In a two dimensional matrix a true test data set is com pared against its classification results Test data with
85. nction name timeSynchro m This function synchronizes two accelerometers which experienced one shared event e g a light hit against each other The separating time shift of both sensors is esti mated by the time difference of the peaks caused by the shared event First the func tion asks for two CSV files which recorded the shared event The input of the sensors IDs is mandatory Then the user is asked to mark the peaks of the shared event The function calculates the time shift saves it in a DAT file and plots the time synchronized acceleration data for visual validation Import of two data sets of two different sensors Both data sets are imported with the RAW Conversion m function see appended CD The naming of the sensors should be the hardware serial number or a unique institutional naming assuring an unambiguous identification Consequently naming the sensors with the same ID is not possible http www mathworks com matlabcentral fileexchange 22663 weeknum content weeknum m 33 5 IMPLEMENTATION Calculation of absolute acceleration on all axes To identify the peak of the shared event in the acceleration data the absolute accelera tion on all axes is plotted For its calculation the Euclidian norm is used see chapter 3 2 1 calculate the absolute acceleration m A length dataX_A dataAbs A zeros m_A 1 Eege al An EELER end Choose and save timestamp of peak After plotting the absolute accel
86. nding stairs descending stairs jogging were under consideration Two tri axial accelerometers were attached to the hip and parallel to the tibia Pattern characterizing features from the time domain mean standard deviation absolute maximum and the frequency domain main frequencies spectral entropy autoregressive coefficients signal magni tude area were extracted Artificial neural networks ANN with a feedforward topology were trained with backpropagation as supervised learning algorithm An evaluation of the resulting classifier was conducted with 14 subjects completing an activity protocol and a free chosen course of activities An individual ANN was trained for each subject Accuracies of 87 99 and 71 23 were approached in classifying the activity protocol and the free run respectively Reliabilities of 96 49 and 76 77 were measured These performance parameters represent a working ambulant physical activity monitor ing system Key words Activity recognition accelerometer artificial neural networks ambulatory monitoring supervised learning ABSTRACT Abstract Deutsch Thema der Masterarbeit Mustererkennung menschlicher Bewegungen in Beschleunigungsdaten Autor Erstpr fer Dipl Ing FH Dennis Jos Prof Dr Oliver Kalthoff Zweitpr fer Dr rer nat Uwe Mittag Bei ambulanten Studien ist das Mitwirken des Probanden Compliance von gro er Bedeutung Besonders bei Interventionsstudien die den menschlichen Be
87. nfusion which is why the user manual was not valued as fully understandable Missing steps were pointed out by the conducting researcher and were corrected appropriately The diagrams which present the results of classifications also received a satis factory validation Neither is there any superfluous diagram nor is any unnecessary information shown Furthermore the diagrams allow a simple and quick overview of the results and are therefore very likely to be used in future research work In short it can be said that the usability of the software program supports the researcher in her studies and simplifies the work flow 66 7 DISCUSSION AND CONCLUSION 7 2 Conclusion In this work a software program for human activity pattern recognition was successfully developed With the help of trained ANNs a classification of different activities was possible The ANNs were trained with supervised learning algorithms and the required ground truth data was retrieved from activity course acquisitions A combination of fea tures from the time domain and frequency domain were extracted to describe the ac celeration signal Different graphical visualizations represent the classification results Table 7 1 Requirements and their accomplishments Requirements accomplished through When did the subject do which body Overview graph of daily Reg gt Se Y movement physical activity How long did the subject do
88. nnnnnne nennen 28 4 2 3 Feature calculation EE 29 4 2 4 Activity classification cooonninnnnnccnnnnnccnnnoncnoncnnnnncnnnnnanrnnccnnnnnnnnnnnnnnnnno 30 VI LIST OF CONTENTS Implementation EE 32 5 1 Implementation enironmernt ku 32 5 2 Preparation EE 32 5241 Data impon nirre teie einleiten 32 52 2 lee e WEEN 33 5 3 Training Phase un ek 35 5 3 Data CONVOrSION musico alli ke 35 5 3 2 Dynamic nature classification oooonnnnnnnnccccnnnnnnnnnccnccccccnnncnnnnnanrcccnnnnn 37 5 3 3 Feature extraction time domain analysis ccceeeeeeeeeeeeeeeeeenteeeeeeees 38 5 3 4 Feature extraction frequency domain analysis 38 5 3 5 Neural network training excised enticed A 40 Bid Blassifieationphase inn Eee ea 45 5 4 1 Neural network ClaSsifiCation 44 nnnnnnnennnnnnnnnnnnnnnnnnnnnnnnn 45 5 4 2 ele 46 5 5 Graphical visualization Of resuhte nn 47 5 5 1 Classification DEE leede ge 47 EEN e al Ee 51 5 6 Standalone executable EE 51 Evaluation ass 52 6 1 Trial subject recruttment uk 52 6 2 Test procedure and course Of stuchy ANNE 52 6 2 1 Data acquisition ee 53 6 2 2 ACTIVITY Classes nn aa ER HN 56 62 3 LEE sti a rara 57 6 3 Test data processing soii ida 59 6 3 1 Training and classttcaton 59 6 3 2 Evaluation results He rer a eg 59 EN Description Ol result ls 61 Discussion and CONCIUSION WE 65 TA DISCUSSION EE 65 2 22 GONCIUSIOM rer ee ee 67 13 QM meer een
89. o 2 pp 128 132 2011 D Hristu Varsakelis and W S Levine Handbook of networked and em bedded control systems Boston Birkh user 2005 XIII ITC113 Karr10 KhLKO8 KLLK10 KNEKO8 Kram09 Kraw13 KwWM11 LKK 10 REFERENCES ITC ILWIS Confusion matrix Online Available http spatial analyst net ILWIS htm ilwismen confusion matrix htm Accessed 24 Oct 2013 U Karrenberg Signale Prozesse Systeme eine multimediale und inter aktive Einf hrung in die Signalverarbeitung Heidelberg Springer 2010 A M Khan Y K Lee and T S Kim Accelerometer signal based hu man activity recognition using augmented autoregressive model coeffi cients and artificial neural nets in Engineering in Medicine and Biology Society 2008 EMBS 2008 30th Annual International Conference of the IEEE pp 5172 5175 2008 A M Khan Y K Lee S Lee and T S Kim Accelerometer s position independent physical activity recognition system for long term activity monitoring in the elderly Medical and Biological Engineering and Com puting vol 48 no 12 pp 1271 1279 2010 T Van Kasteren A Noulas G Englebienne and B Kr se Accurate activity recognition in a home setting in Proceedings of the 10th Inter national Conference on Ubiquitous Computing pp 1 9 2008 O Kramer Computational Intelligence Eine Einf hrung Berlin Heidel berg Spring
90. occurred with Sitting and Standing This indicates difficulties for the classifier to distinguish between these two activities This confusion seems reasonable because the position of the hip may not change noticea bly Since Sitting and Standing are static activities the orientation of the accelerometer is a decisive factor for the classification The lowest producer s accuracy was measured for the activity Walking slow Apparently walking slowly is a type of locomotion where the acceleration patterns vary substantially It seems that the instructions for walking slowly mast have been vague so the subjects were unable to repeat this locomotion in the same manner Furthermore the walking behavior may differ in various surroundings Walking slow may be per formed differently on a big parking lot than in a small corridor However a higher relia bility of Walking slow indicates that the classifier is able to distinguish it from other lo comotion The remaining reliabilities exceed 80 which is a satisfying result The same conclusions are applicable to the results of the tibia sensor during the AC classification Looking at the performance parameters of the classifier it seems consistent that accuracy and reliability were high for the AC classification because environment pace and motivation were identical for training and test data set Nevertheless this fact does not prove the proper functionality of the analyzing algorithm in the recognition
91. oftware program vi APPENDIX DEUTSCHES ZENTRUM FUR LUFT UND RAUMFAHRT E V DLR Movement Identification Tool MovelT User Manual uwe mittag dlr de Deutsches Zentrum f r Luft und Raumfahrt e V DLR German Aerospace Center Institut f r Luft und Raumfahrtmedizin Institute of Aerospace Medicine Abteilung Weltraumphysiologie Department of Space Physiology Linder H he Linder H he 51147 K ln 51147 Cologne Deutschland Germany vii 1 2 3 4 5 6 7 APPENDIX Content a core Bes de es ee a RO ix et ALON Scoop a ee eat a Ml ites ok Eee ix 2 1 Working on a computer with MATLAB license ooooooccccnnnccccnncanoncccnnnncnnnnnnnnnncccnnnnnos D 2 2 Working on a computer without MATLABO license nn xi Configuration of accelerometers EEN xi 3 1 Howto use the XLR8R sottiware ENNEN xii 3 2 Mandatory settings of accelerometers for the use with MovelT nooo xii Getting started with Movell a os chccteeeranen a xiii 4 1 How to set the default folder path AAA xiii 4 2 Howto import data xiv Working with tibia and hip sensorg n XV 5 1 How to synchronize two accelerometers un XV 5 2 Howto train a neural network ENEE xvi 5 3 HOw 10 classify activities 2 xvii 5 4 How to check for subjects compliance AAA xix Working with a single sensor u xix te EE XX Til A ee Reel XX viii APPENDIX 1 Introduction The Movement Identification Tool MovelT is an easy to use software program for ana
92. orm A fast walking is to be avoided 56 6 EVALUATION 6 2 3 Activity course All the activities described above were performed and recorded during an activity course see Table 6 2 The course consisted of different stations each demanding a certain movement Between stations the sensors kept recording accelerations These movements could not be allocated precisely to an activity but still needed to be proto coled as interstation activities For better differentiation of the activities in the visual representation it was im portant to start and end dynamic activities with static ones and vice versa According to the durations stated in Table 6 2 one run through the course took less than 15 minutes Every subject had to complete the course twice Afterwards the subject was ordered to do a third individual run The subject was free to go around the building moving with any type of locomotion The only requirement was that every activity from Table 6 1 was performed for at least 10 seconds throughout the entire run Figure 6 4 shows a map of the DLR campus and where the different activity stations were located az Deutsches Zentrum DLR f r Luft und Raumfahrt German Aerospace Center ia A Basement tunnel K ln Porz e 1 standing J gpd 4 walking normal 5 jogging distance for walking slow N Staircase 2 ascending stairs 3 descending stairs Ground floor 6 walking slow Physio
93. overfitting In this case the classifier is able to classify the training data correctly but new similar test sets tend to be misclassified Its result is the unintended inability to generalize Kram09 Hidden Input Output Figure 3 1 Artificial neural network topology One of the most common learning algorithms is backpropagation This method is based on supervised learning and requires a certain network topology It usually consists of an input layer an output layer and hidden layers between them see Figure 3 1 Units in the input layer receive the input features The output layer units provide the response of the ANN to the input data The connections between units are directed to the next higher layer and not recurrent from higher to lower layer This kind of net work is also known as feedforward network The number of units per layer is depen dent on the feature dimension and the output classes This type of ANN is preferred for prediction and classification tasks Kram09 AgBe00 16 3 MATERIALS AND METHODS The basic idea of backpropagation is minimizing the percentage of misclassifi cations Learning means reducing the error by manipulating the connection weightings The name of this method refers to the way the error computed at the output layer is propagated backwards to the lower layers In the first of two phases the forward pass initial values for the weightings are assumed The input is propagated forward through
94. pe of running to be measured Looking at the amplitude range of several subjects lightly jogging the accelerations at the tibia position did not exceed the range of 6 g to 6 g In conclusion the accelerometers are set to a sample rate of 20 Hz and a low gain so the detecting ranges from 6 g to 6 g These settings allow assessing physi cal activity and additionally prolonging battery life Conversion The data measured by the accelerometers is saved as raw count data Its conversion into acceleration data is inevitable Gain settings change the proportionality coefficient between count data and g force units see chapter 3 4 To make the software program applicable for unknown future use the conversion process is designed to choose the corresponding proportionality coefficient according to the received data Signal noise can affect the outcome of the analysis Such interference can be lowered by filtering the data before the analysis Literature review has shown that sev eral studies have applied different filter methods during pre processing acceleration data HANSON ET AL HGP 11 used two accelerometers simultaneously sampled the data with 256 Hz waist sensor and 1024 Hz foot sensor and processed each data set with a low pass filter at 100 Hz A sample rate of 2000 Hz and filtering by a fourth order low pass filter with a cut off frequency of 45 Hz was chosen by GENC ET 27 4 DESIGN CONCEPT AL GeMC06 Due to these chosen s
95. ponded to the first one but its data was intended for testing of the ANNs The order of a third run was chosen individ ually by the subject and served as test data as well 52 6 EVALUATION 6 2 1 Data acquisition Main aspect of this evaluation was the acquisition of acceleration data from laboratory recordings Accelerometers from GCDC were used as sensors Subjects received two accelerometers one on the lower leg and one on the hip The lower leg sensor was placed at a similar position as it was during the HEP study WDM 13 see Figure 6 1 a a b Figure 6 1 Position of accelerometer on HEPHAISTOS orthosis a and on shin guard b It was not possible to run the evaluation with subjects wearing a HEP orthosis because it is a custom made orthosis Each subject would have needed an expensive custom built model These costs over 60 000 per orthosis were not reasonable for this eval uation Moreover the subjects from the past HEP study were not available Instead gear for mounting the accelerometer to the lower leg was used A prototype was creat ed using a shin guard as it is common in soccer sports see Figure 6 1 b The sensor was placed parallel to the tibia bone and secured with one screw to the shin guard Adhesive tape was added to avoid a tilting of the sensor The placement of the sensor with a shin guard was an adequate substitution to the HEP orthosis 53 6 EVALUATION Figure 6 2 Belt bag with
96. probability of error However to take observed data into account independence assumptions have to be made which tend to be inaccurate DeSS11 LoYA09 Decision tree classifier A decision tree consists of internal nodes branches and leaves Every internal node represents an attribute or feature The branches are a test on the node from which they are coming from Every leaf represents an output class Decision trees are con structed according to a training set For this purpose there are different algorithms available A decision tree is applied to new data by running through the tree from the root node to a leaf The ending leaf corresponds to the classification result Decision trees describe relationships between features and output class by using simple deci sion rules Therefore they are easy to understand for the user B hm03 Cimi07 A decision tree classifier is a decision tree used for classification Numerous learning methods have been proposed But most of them have a tree growing and pruning phase in common Dobr09 14 3 MATERIALS AND METHODS Cluster analysis Cluster analysis is an approach to find unknown classes or clusters in a set of data objects Every object has a number of features Clusters are found where features of objects in the same cluster are as similar as possible and differentiate as much as pos sible from objects in other clusters There are numerous clustering algorithms available for structuring
97. problems the developer will interfere After completing several tasks the researcher is asked to fill out a usability questionnaire see appendix 9 1 This questionnaire was designed to give a short overview of the impressions of the software to the user The results can be sighted in appendix 9 1 and are not described separately However the discussion of the results will be covered in chapter 7 1 20 3 MATERIALS AND METHODS 3 6 State ofthe art This chapter describes the state of the art in human activity pattern recognition from acceleration data An approach of activity recognition in a home environment with inexpensive an notation tools and easy installation was developed by VAN KASTEREN ET AL KNEKO8 In their work activities like showering breakfast dinner or sleeping were under con sideration For this digital state change sensors were installed in a subject s home A Bluetooth headset was worn by the subject to annotate the current activity thus gen erating a ground truth data set with known output Raw point of change and last observation data was classified with hidden Markov models and conditional random fields VAN KASTEREN ET AL achieved the best reliability of 95 6 and the highest ac curacy of 79 4 with conditional random fields and hidden Markov models respectively CHUNG ET AL ChPS08 focused in their publication on a real time application of activity recognition With wireless MEMS accelerometers
98. program 2 REQUIREMENTS A numerical computing environment is needed to implement these procedures preferably one with built in functions for signal processing and the chosen classification method Interpreting Req3 The stated requirement is a two class classification problem The classification meth ods described above may also be applicable in this scenario Finding a simple decision rule which only distinguishes between compliance and non compliance is however less complex Further explanation of this decision process will be given in chapter 4 2 4 Interpreting Req4 to Req9 Requirements Reg4 to Req9 are directed to the user interface and the application of the software program and have no effect on the analyzing algorithm Usage related requirements Req4 Req5 Req7 and Req8 are to be concerned when designing the graphical user interface GUI Compiling software is necessary to compile a standalone executable Req6 Interpreting Req10 Like with most mechanical sensors accelerometers measure the impact they are ex posed to by changing voltage These changes need to be converted into acceleration data using specific calibration equations which are to be investigated Further explana tion of the conversion process will be given in chapter 3 4 In conclusion the following system requirements arise from the user requirements SysReq1 Classification method for pattern recognition SysReq2 Feature extraction methods for signal
99. r analysis is the readability of the results They are not easily comprehensible and need some kind of interpretation HeKi08 Further more it is an unsupervised learning algorithm which according to WAGSTAFF uses very general notions to identify patterns and interesting structure in data Wags02 Unsupervised learning algorithms are unguided and may tend to focus on uninterest ing patterns e g patterns due to systemic errors A supervised learning algorithm is more suitable to the problem of human activity pattern recognition because known activities are to classify ANNs grew from basic research to establishing practical implementations of in novative applications Karr10 Where linear models are unable to describe a model accurately and the underlying relationships are not well known a nonlinear neural net 23 4 DESIGN CONCEPT work is helpful HrLe05 Tu96 ANNs have a proven record of success in human acti vity recognition HGP 11 KLLK10 KwWM1 1 LKK 10 Additionally they are easy to conceptualize because numerous libraries and implementations are available How ever an ANN works as black box because connection weightings of a trained ANN do not reveal relationships between input features But if the classification result is in fo cus this issue is considered as reasonable compromise ANNs were often blamed for randomness in results and being unpredictable The latter accusation is without any reason be
100. ram During the work on this thesis the conducting researcher of the NutriHEP study decided to integrate skin conductance sensors in the HEP orthosis These sensors can detect skin contact and thus if the orthosis is worn or not This technology promises a high accuracy Therefore the requirement for the compliance check was downgraded and seen as supporting feature Consequently the results of the compliance check are not under consideration in this evaluation 6 1 Trial subject recruitment Subjects were mainly recruited from the staff of the Department of Space Physiology A monetary compensation was not obtained The number of subjects was 14 and includ ed 7 female and 7 male subjects so any gender variances in movement due to anato my were considered The average age was 29 4 9 3 years mean SD For each subject the study lasted for approximately one hour of laboratory data acquisition Ex clusion criteria were abnormalities of gait or mobility rheumatic diseases and preg nancy The study contained activity measurements in the physiology lab and the build ing of the nstitute of Aerospace Medicine The subjects were not dictated to any spe cial diet or medication 6 2 Test procedure and course of study The subjects wore accelerometers and ran through a given course where several ac tivities were performed The first run had a predefined order and time plan and served as training data for the ANNs The second run corres
101. re saved in separate matrices In addition the time stamps of the dynamic nature are saved as DynamicNatureTimes m This file con tains the points of time where the signal was classified as static or dynamic This file is relevant for the compliance check in the following chapter 5 4 2 Compliance check Function name complianceCheck m This function is used for checking the subject s compliance during the NutriHEP study It compares the dynamic nature of the acceleration signal measured at hip and tibia If a Static activity is detected on the tibia but a dynamic one on the hip then it is consid ered that the subject is not wearing the tibia sensor thus the orthosis In all other cases all sensors are assumed to be worn At the end the result is plotted in a graphic Import classification results The results from the classification are loaded from the file DynamicNatureTimes m It contains time stamps separated in static and dynamic activities First the time stamps are concatenated into one matrix for each sensor position hip and tibia After this the matrices are sorted in time which results in two matrices hip and tibia with time stamps and dynamic nature Static activities are encoded as 0 and dynamic acti vities as 1 o load dynamic nature classifications results CCtibia load fullfile CCfolder tibia DynamicNatureTimes mat CChip load fullfile CC folder hip DynamicNatureTimes mat init and
102. rlap percentage and activity template This master s thesis presents a human activity pattern recognition system which will be used in the NutriHEP study The software program will monitor the activity distribution of a subject and check that the orthosis was worn A further possible appli cation is during bed rest studies where it would be interesting to examine a subject s activity profile before and after bed rest Moreover the developed software program may be useful in the field of ambient assisted living where activity recognition is an important issue for the situation adaptability of home care systems 69 REFERENCES 8 References AgBe00 Alve07 BaHa00 Beuc11 BKV 97 BMSG80 Bohm03 ChPS08 S Agatonovic Kustrin and R Beresford Basic concepts of artificial neu ral network ANN modeling and its application in pharmaceutical re search Journal of Pharmaceutical and Biomedical Analysis vol 22 no 5 pp 717 727 2000 T M Alves Spektrale Entropie und Bispektral Index als Messgr en f r die Wirkung von Propofol auf das EEG Universit ts und Landesbiblio thek Bonn 2007 A Basheer and M Hajmeer Artificial neural networks fundamentals computing design and application Journal of Microbiological Methods vol 43 no 1 pp 3 31 2000 O Beucher LTI Systeme und Stochastische Signale in Signale und Systeme Theorie Simulation Anwendung Berli
103. rule is applied for the classification of the dynamic nature of human activity WDY 13 4 2 3 Feature calculation Static activity classification For the entire time phase where the activity is of static nature time domain features are extracted as described in chapter 3 2 1 Frequency domain features are not applicable in this case Descriptive methods like mean SD and absolute maximum see Table 4 1 are considered as traditional features that are used for acceleration activity recognition GjGC10 and have been successfully applied by HANSON ET AL HGP 1 1 PREECE ET AL PGKHO9 BOUTEN ET AL BKV 97 and many more Dynamic activity classification For the classification of dynamic activities time and frequency domain features are extracted The peaks of the power magnitude represent the main frequencies and are quantified by the first components of the FFT analysis It seems reasonable to investi gate the frequencies in activity patterns because different types of locomotion underlie different frequency patterns PREECE ET AL PGKHO9 showed in their work that the magnitude of the first five components of an FFT analysis have the best classification accuracies with sensors worn on waist and ankle However in this thesis the first three components are used as describing feature Furthermore the periodicity was a promising feature for Hanson et al HGP 11 and Lee et al LoY A09 SE is a feature which was mainly introd
104. s of the software and their algorithms and basic functions are explained in more detail 5 1 Implementation environment The software development uses the following external software e MATLAB R2012b v8 0 0 783 The MathWorks Inc Natick MA USA e Neural Network Toolbox The MathWorks Inc Natick MA USA e Signal Processing Toolbox The MathWorks Inc Natick MA USA e MATLAB Compiler The MathWorks Inc Natick MA USA e USB accelerometers X6 1A and X6 2 Gulf Coast Data Concepts Waveland MS USA 5 2 Preparation phase This chapter explains the implemented code in MATLAB for the software functions which are necessary beforehand Only fundamental code lines are presented here The full program code is accessible on the appended CD 5 2 1 Data import Function name retrieveData m This function allows automatic storing of acceleration data from a USB drive to a hard drive The GUI language is chosen by the calling main function The user selects the USB root drive Its details and header information are retrieved The function automati cally sorts the CSV files in the following folder hierarchy subject ID calendar week and sensor position This structure makes it easier to compare activities between dif ferent weeks The files are renamed with date and time of starting acquisition This avoids overwriting of different data sets with the same naming Read out USB Sensor details from configuration file and header The
105. s one activity Therefore the window size corresponds to the resolution of activity classifica tion For example a window with a length of 10 seconds allows the classification of an activity every 10 seconds This is the case if the next window is consecutive The reso lution can be increased by using overlapping moving windows 3 2 Feature extraction methods During the process of feature extraction several parameters are calculated from the acceleration signal representing the characteristics of an activity and thus making them identifiable This section illustrates their basics taken from the fields of descrip tive statistics and Fourier analysis 3 2 1 Descriptive statistics Mean value The mean value describes the average value of a digital signal Measuring with tri axial accelerometers the individual means of all three axes in combination represent the orientation of the sensor towards the gravitational field of the earth It is easy to calcu late and an important indicator for body orientation Standard deviation The standard deviation SD is a measure of dispersion from the signal s mean value It appears plausible that more fluctuation of the signal corresponds with more activity of the sensor and consequently of the subject It can serve as indicator for the dynamic nature of a signal The method for classifying the dynamic nature will be described in chapter 4 2 2 Absolute maximum The maximum value refers to
106. signal by hundreds Smit03 A divide and conquer application leads to a decreasing number of required op erations M ll13 According to the Radix 2 FFT algorithm by COOLEY AND TUKEY CoTu65 several symmetric properties allow a numerical efficient calculation of the DFT Neub12 Implementations of the FFT allow signal analysis in the frequency do main without higher computational expense In MATLAB the FFT functions are based on the FFTW library FrJo98 using the Cooley Tukey algorithm CoTu65 to compute an N point DFT N needs to be com posite i e N N N2 First N transforms of size N are computed and then N transforms of size N The decomposition is applied recursively until the problem can be solved using one of several machine generated fixed size codes These codes in clude combinations of the Cooley Tukey algorithm OpSc10 a prime factor algorithm OpSc10 and a split radix algorithm DuVe90 The computation time of the MATLAB FFT function depends on the length of the transform Math13 The FFT is a basic tool in signal analysis and an essential part of activity recog nition It is written as Y f FFT x t Equation 3 5 adapted from Math1 3a 11 3 MATERIALS AND METHODS Periodogram As stated in Equation 3 5 Y f gives the distribution of the Fourier coefficients in the complex plane The power of the input signal is the squared complex magnitude of Y f see Equation 3 6 Plotting it against its sca
107. ss corridor stop and stand still at physiology lab door 30s Interstation activity stand still for 15 seconds walk into physiology lab walk to chair and stand back 60s Physiolo gy Lab 13 00 Sitting sit down on chair both feet touch ground 30s Interstation activity get up from chair walk to couch and stand back 30s 14 00 Lying lie down on couch lie on back do not cross legs 30s Interstation activity get up from couch walk back to chair turn accelerometers OFF 30s XX
108. the text field 2 Type in your folder path of choice The program checks for the existence of the chosen folder A default folder path has to be chosen to run any of the following functions Please note At start of MovelT the default folder path is always set to G AWP STD NutriHEP NHP b Durchf hrung Accelerometerdaten A change of the path is only saved until the home screen window of Move IT is closed xiii APPENDIX 4 2 How to import data MovelT contains a feature for retrieving data from an USB accelerometer to your hard drive or network drive 1 2 Ego Under Preparation Phase press Import Data Choose the USB drive of the accelerometer with the data you want to copy and press Ordner ausw hlen Subject ID sensor position and date of acquisition is shown for verification Choose whether you want to save it under the default folder path or a new path The data is saved with the following folder structure if default folder path cho sen e Default folder path e Subject ID e Calendar week e Sensor position The loading process can be canceled by closing the progress window xiv APPENDIX 5 Working with tibia and hip sensors The development of MovelT was conducted in consideration of the upcoming NutriHEP study at the German Aerospace Center This study will monitor each subject via two accelerometers placed at hip and tibia This issue is faced in MovelT in particular The following documentation will guid
109. tivities from the activity catalog are listed on the left For each activity the number of successful recognitions its percent age share and the estimated share of time is displayed The calculation of the per centage share is equal to the algorithm for the pie chart The estimated time of each activity is computed by taking the respective percentage of the total elapsed time Table 5 1 Table results GUI Days Hours Minutes Seconds 00 00 14 59 565 Avg certainty of Classification dynamic activities 99 7922 static activities 95 7240 all activities 97 5903 No of recognitions Activity distribution estimated Time dd hh mmiss unclassifiable 10 3 3003 00 00 00 29 standing 125 41 2541 00 00 06 11 sitting 6 1 9802 00 00 00 17 lying 8 2 6403 00 00 00 23 walking normal 88 29 0429 00 00 04 21 walking slow 9 2 9703 00 00 00 26 stairs up 28 9 2409 00 00 01 23 stairs down 6 9307 00 00 01 02 jogging 2 6403 00 00 00 23 sprinting 0 00 00 00 00 SUM 100 00 00 14 59 This graphical visualization of the results is mainly intended for looking at a quantitative activity distribution profile It is plotted with the function plotTable m 50 5 IMPLEMENTATION 5 5 2 Compliance results The compliance is represented in a colored timeline Timeframes with worn and not worn tibia sensor are encoded in green and red respectively Compliance check of Tibi
110. tput uncorrelated error p yO X ayt D e i 1 Equation 3 10 KhLKO8 The number of past values y t i which were used to estimate the current value for y t defines the order p of the AR model KhLK08 The AR coefficient can be estimated with the Burg method This method esti mates the reflection coefficients and uses them to estimate the AR coefficients recur sively Math13c DORFFNER evaluates AR coefficients as not exact descriptive but sufficient esti mation The computation is rated as efficient Dorf13 13 3 MATERIALS AND METHODS Signal magnitude area The signal magnitude area SMA is defined as the area under the magnitude of the root mean square of all three axes ChPS08 It is calculated as N sMA Y xOD OD D 7 a KLLK10 where x i y i and z i represent the acceleration signal along x y and z axis respectively 3 3 Classification methods Classification is the allocating of elements to classes according to their characteristics These characteristics are called features In principle elements with similar features belong to the same class Kram09 Bayes classifier A Bayes classifier is based on the Bayes theorem and used for probabilistic learning lt requires a probabilistic model or cost function which estimates all misclassifications or unclassifiable data Correct classifications do not occasion any costs By minimizing these costs a Bayes classifier has the lowest possible
111. ts from the input signal Incrementing the av eraging number leads to a decreased noise in the signal However the step response is lowered It is mainly used for its smoothing effect Low pass filters process an input signal by passing only low frequency components They reduce the amplitude of high frequency signals above a specific cutoff frequency Hence low pass filters are used to reduce high frequency noise Band pass filters pass frequencies in a certain range Amplitudes of signals with frequencies outside this range will be reduced They can be incorporated by combining a low pass and high pass filter simultaneously Therefore a band pass filter has two cutoff frequencies The filtering methods above are the basis of signal separation and restoration and have been applied by HANSON ET AL HGP 11 KHAN ET AL KLLK10 LONG ET AL LoYA09 and many more Numerous implementations of different filtering methods are available and easy to include but due to the manipulating effect the possibility of los ing important information from the signal is present 3 MATERIALS AND METHODS Moving Window To extract several features from a signal stream the data needs to be divided by small windows This ensures an efficient computation time considering the idea of the divide and conquer paradigm Additionally it is inevitable for the classification of activities A window which contains a certain number of data samples can be identified a
112. uced by KHAN ET AL KLLK10 LKK 10 They claimed it to be their best discriminating feature in recognizing resting upper body and lower body activity ERMES ET AL ErPC08 used SE for differentiation between running or walking and cycling KHAN ET AL KLLK10 also used the AR coefficients and the SMA as describing features and produced promising results In a preceding work KHAN ET AL KhLK08 found with autocorrelation analysis that an AR model of order 3 is most suitable Addi tionally CHANG ET AL ChPS08 found different activities like running and walking have different SMA levels Whereas KHAN ET AL KLLK10 significantly improved the recognition rate of dynamic activities with a combination of SMA with AR coefficients It 29 4 DESIGN CONCEPT was the best discriminating feature for all activity classes and all sensor positions A proportional correlation between SMA and energy expenditure was approved by MATHIE ET AL MCLC04 and BOUTEN ET AL BKV 97 Table 4 1 Feature list Domain Feature Classification of Mean Time gt Static and dynamic Standard deviation ER domain activities Absolute maximum First 3 peaks in power magnitude Frequency Spectral entropy Dynamic activities domain Autoregressive coefficient Signal magnitude area 4 2 4 Activity classification Classification with neural network The hierarchical order of multiple ANNs is based on the structur
113. vacuuming Again a ground truth data set was ensured by having the subjects use a Bluetooth headset with speech recognition SE AR coefficients and 21 3 MATERIALS AND METHODS SMA were the features of choice A two level recognition approach included ANNs and linear discriminant analysis An accuracy of 90 was achieved LEE ET AL LKK 10 established a real time activity recognition system which would serve as personal log life system Every day activities were distinguished includ ing driving and climbing stairs Rather impractical non wireless accelerometers were attached to the chest To generate a ground truth data set camera recordings were analyzed Classification features were SD SE correlation among all axes AR coeffi cients SMA and tilt angle Furthermore stride length and step count were extracted for calculating distance speed and energy expenditure LEE ET AL constructed a two level classifier which first distinguishes between static and dynamic activities On the sec ond level two ANNs static and dynamic were trained with backpropagation algorithm An accuracy of 84 8 was achieved In the NASA Glenn Research Center HANSON ET AL HGP 11 worked on a new method for tracking crewmember activity during space missions in reduced gravity environment They employed an enhanced zero gravity locomotion simulator to meas ure different types of locomotion under different gravitational influences Two wireless accelero
114. variables activityClass and Nstatic are declared as global varia bles Their validity for all sub functions is essential for adding new activities to the cata log The activity catalog contains the names of all activities which are to be classified It is saved as a list in activityClass where the first elements are of static nature If Nstatic is 3 then the first three elements in the activity catalog are static activities The implementation of a variable activity catalog is essential for the unknown use with a single accelerometer global activityClass global Nstatic Set standardized activity catalog SHEET ySls Standing sreeing r Eeer walking normal walking slow Ve bane EE musica isn SiGe Mas er Jogging sprankeingl Nstatic 3 Select data The user selects a folder with training data set for hip and tibia position via a browsing window The subject ID is read out from the folder path A new function train dynNN m is called to start the training for tibia and hip data individually The trained ANNs are saved in the training folder Function name train_dynNN m This function trains an ANN with training data which is already allocated to activities Acceleration data is imported and its dynamic nature is classified For the training of the ANN the feature extraction uses a window size of 6 seconds For the actual train ing and testing of the ANN a sub function generated with MATLAB is used trainNN m
115. ware should work with GCDC accelerometers GCDC accelerometers are sensors manufactured by Gulf Coast Data Concepts Waveland MS USA These accelerometers were used during the HEP study a pre vious study run to develop the HEP orthosis The NutriHEP study will be working with these accelerometers as well Therefore the concept of the analysis software should be focused on the processing of data acquired with GCDC accelerometers In addition the software might offer the possibility for a post hoc analysis of the HEP data 2 REQUIREMENTS 2 2 System requirements Interpreting the previous user requirements various system requirements SysReq for the development of this software program arise The analysis of acceleration data re quires numerous procedures and methods Interpreting Req1 and Req2 Concluding from acceleration data to body movement is a common classification prob lem It needs to be identified to which category a new observation belongs From a plethora of classification methods available the following ones occur in many publica tions Bayes classifier LoYA09 decision tree classifier LoYA09 and artificial neural networks HGP 11 KLLK10 Further explanation of these methods will be given in chapter 3 3 Classification algorithms cannot process raw signal data Instead they find cor relations and similarities in features Features are properties of raw signals which allow representing the characteristics of
116. wegungsap parat untersuchen ist es wichtig die allt glichen Aktivit ten in heimischer Umgebung zu protokollieren K rperliche Aktivit t hat Einfluss auf Messungen die Auswirkungen von Ern hrung Stoffwechsel oder neuromuskul ren Stimulationen quantifizieren Zur Unterst tzung einer ambulanten Studie am Deutschen Zentrum f r Luft und Raum fahrt DLR wurde ein System zur Erkennung menschlicher Bewegungen entwickelt Allt gliche Aktivit ten von statischer Stehen Sitzen Laufen und dynamischer Natur Gehen Treppen auf bzw absteigen Joggen wurden untersucht Zwei dreiachsige Beschleunigungssensoren wurden an der H fte und parallel zum Schienbein montiert Von den Beschleunigungssignalen wurden musterbeschreibende Merkmale aus dem Zeitbereich Mittelwert Standardabweichung Betragsmaximum und Frequenzbereich Hauptfrequenzen spektrale Entropie autoregressive Koeffizienten Fl che der Signal amplituden extrahiert K nstliche neuronale Netze ANN mit einer Feedforward Struktur wurden mit Backpropagation und berwachtem Lernen trainiert F r die Eva luation des daraus entstandenen Klassifikators absolvierten 14 Testpersonen einen Aktivit tsparcours und einen freien Lauf mit beliebiger Reihenfolge von Aktivit ten F r jeden Probanden wurde ein individuelles ANN trainiert Genauigkeiten von 87 99 und 71 23 wurden bei der Klassifikation des Aktivitatsparcours bzw des freien Laufs gemessen Die Zuverl ssigkeit der Klassifiz
117. which body Overview graph of daily Req2 o i n Y movement physical activity Req3 Was the orthosis worn the entire time Overview of compliance V Read What does the data look like during uni Overview graph of daily M aq dentified activities physical activity Req5 The software should be easy to use Usability test m The software should work as a SN Req6 EES SE Standalone application Y Fear 7 Soma SOULS provided IEP Ii uilingualinplementatlen E lish and German Reqs The Saale should not be time con Usability test Li suming Req9 a Giel ECH WAR einge Successful implementation m sensor on any position The software should work with GCDC Use of GCDC accelerome Req10 vi accelerometers ters Compliance can be checked but the results are unevaluated In chapter 2 numerous user requirements were collected which have been all accomplished through one software program see Table 7 1 Diagrams showing the classified activity over any given time can be generated Req1 The user can also have a look at the raw acceleration data and review unidentified activities Req4 Pie charts show the distribution of activities Req2 A compliance check can be executed to see if the orthosis was worn or not Req3 The ease of use of the GUI was tested 67 7 DISCUSSION AND CONCLUSION and approved in a usability test Read It offers an English and German language package Req7 and
118. with pre trained neural networks Therefore the feature to train a neural net work is disabled Installation of MCR MCR stands for MATLAB Compiler Runtime and allows the execution of compiled MATLAB applications on computers with no MATLABO license 1 Go on the DLR internal server to G WP IT MovelT 2 Copy the folder Standalone Software to your local hard disk drive 3 Run the file MCRinstaller exe Please note You will need administrator rights to run this file Ask your departmental IT administrator for further help The MCRinstaller exe is also available online on http www mathworks com products compiler mcr Installation of MovelT The standalone application of the MovelT software does not need further installation Copying the executable file is enough 1 Go on the DLR internal server to G WP IT MovelT 2 Copy the folder MovelT Software exe to your local hard disk drive Starting MovelT After successful installation of the MCR the standalone file can be executed 1 Run the file MovelT Software exe 2 The MovelT Software will start showing the home menu See Getting started with MovelT on page xiii 3 Configuration of accelerometers Working with MovelT requires certain configurations of the accelerometers The follow ing instructions refer to accelerometry sensors from Gulf Coast Data Concepts GCDC Onboard software for configuration is included Please note For now MovelT only works with accelerom
119. y video sighting This implied a subjective interpretation of types of locomotion The diffi 60 6 EVALUATION culty was to judge unclassifiable movements performed by the subjects Most common was the case of transition from one movement to another which was often classified as unclassifiable lt seemed correct to evaluate these classifications as successful Longer episodes of one movement were clearly identifiable in the video Movements which were not listed in the activity list in Table 6 1 were allocated as unclassifiable activities For every subject there were 4 confusion matrices created Two for the AC classification with hip and tibia sensor and two for the free run classification with hip and tibia sensor This leads to a total number of 56 confusion matrices 28 of AC and 28 of FR classification To evaluate the recognition software with its ANN as classifying system accuracies and reliability of AC and FR classification were obtained by merg ing all subjects results 6 4 Description of results Table 6 4 shows a summarizing confusion matrix for the hip sensor of the AC classifi cations from all subjects In addition to the absolute number of classifications each cell contains the percentage relative to the total number of classifications i e in Table 6 4 in the first cell 129 correct classifications of Standing represent 7 13 of all classifica tions The highest occurrences of misclassifications off diagonal
120. y is a helpful parameter It can be frequency or magnitude independent and allows comparing signals of different complexness Several methods are available for the cal culation of entropy To determine the SE a frequency dependent approach is neces sary Alve07 12 3 MATERIALS AND METHODS ERMES ET AL ErPC08 define SE of an acceleration signal for a frequency band I f2 as ye s PED log P f Equation 3 9 SEG fa o BND adapted from ErPC08 where P f represents the PSD value of the frequency f The PSD values are normal ized resulting in a sum of one in the band f f2 The number of frequency compo nents in the corresponding band in the PSD is N f fa In their work ERMES ET AL ErPC08 used the SE for differentiation between running or walking and cycling KHAN ET AL KLLK10 claim SE to be their best discrim inating feature in recognizing resting upper body and lower body activity 3 2 3 Miscellaneous features Autoregressive coefficient In signal processing an autoregressive AR model is a representation form of a time series signal It is an estimation and characterization of how the output variables of a signal depend on its own previous values In addition it has a direct link to the spec trum of the signal Dorf13 Equation 3 10 defines the AR model of a random process y t in the discrete time t where a az a are the coefficients of the model p the order of the model and e t the ou

Download Pdf Manuals

Related Search

Dokument_1. dokument 1 word dokument 1 microsoft word

Related Contents

Manual de Instalação 7 - Ford Cargo Manual de Instalação 7  Samsung AC018HBMDKE Manuel de l'utilisateur  Webcast Troubleshooting Guide • The preferred browser is Internet  Oracle® Solaris Cluster 4.3 Licensing Information User Manual  MANUAL DE INSTRUÇÕES    Sensor de potencia W.I.N.D. Manual del Usuario      IOMSUNPRXX06-DE 01.indd  

Copyright © All rights reserved. Failed to retrieve file