Proyecto de aula

The training datasets and testing datasets we used in this study are all gen- erated in lab settings. Considering eventually the algorithms will be applied to recognize walk from full-natural activities, we need to further evaluate the models on datasets collected from subjects moving around doing their daily activities unsupervised and unrestricted.

Figure 9.1 shows a passage of daily activity from one of the developers collected using an Android app called Accelerometer Analyzer with phone positioned around waist. Green line indicates acceleration magnitudes. Blue line is the step plot of classification results, where value of 0 indicates non- walk class, and value of 50 indicates walk class. Fifty is an arbitrary number chosen in order to conveniently overlay the prediction results on acceleration. Text boxes on the top of the figure shows activities the subject was doing and their durations. The classification results look promising as the model does not mis-classify some common daily activities, such as cleaning the room and cooking, as good walk.

Figure 9.2 shows a passage of daily activity from one of the developers col- lected using the passive monitor Android app we developed with the phone positioned in shorts pocket. Green line indicates acceleration magnitudes. Blue line is the step plot of classification results, where value of 0 indi- cates stationary class, value of 30 indicates stairs class and value of 50 indi- cates walk class. The subject performed activities such as driving, standing, sitting, walking and stairs for over two hours. The passive monitor filters out stationary activities by applying standard deviation threshold, and push moving activities to the model to further check whether activity is walk.

A detailed documentation of durations and types of activities performed in Figure 9.2 is not available. Out of 198 10-second sessions, 34 of which are classified as stationary, 110 as stairs and 54 as walk. The model classifies the majority of activities as stairs class, however up/down stairs only account for a small portion of actual activities performed. The reason may be that the walks performed are not considered as good enough by the model. As shown in the previous example (Figure 9.1), activities such as cooking and cleaning, which are partially walking around, are not recognized as walk class. Since those activities clearly belong to moving class and do not qualify as good

walk, they are recognized as stairs. Low recall on stairs recognition is not a major concern, because for our application we only need to ensure that the model can recognize reasonable amount of good walk with high precision.

The good-walk recognition model used in both applications is standard deviation threshold method and stairs vs. walk random forest combined (Algorithm 1). Given features extracted from an activity session, we first compare the standard deviation with selected threshold (0.7), if it is less than the threshold, the session is classified as stationary. Otherwise we query the stairs vs. walk random forest model for classification result.

Algorithm 1 Standard Deviation Threshold and Learning Algorithm Mixed Model for Good Walk Recognition

procedure WalkClassification data ← activity session featuress

std ← activity session acceleration magnitude standard deviation threshold ← 0.7

clf ← Walk vs. Stairs random forest classifier if std < threshold then return Stationary else return clf.predict(data) end if end procedure

Figure 9.1: Good Walk Recognition on Daily Activities

Figure 9.2: Good Walk Recognition on Daily Activities Collected from Passive Monitor

REFERENCES

[1] A. Smith, “Smartphone ownership2013 update,” Pew Research Center: Washington DC, vol. 12, p. 2013, 2013.

[2] J. Sarasohn-Kahn, How smartphones are changing health care for con- sumers and providers. California HealthCare Foundation, 2010. [3] O. IDC-Smartphone, “market share 2014,” 2015.

[4] S. Fox and M. Duggan, “Mobile health 2012,” Washington, DC: Pew Internet & American Life Project, 2012.

[5] R. A. Rabinovich, Z. Louvaris, Y. Raste, D. Langer, H. Van Remoortel, S. Giavedoni, C. Burtin, E. M. Regueiro, I. Vogiatzis, N. S. Hopkinson et al., “Validity of physical activity monitors during daily life in patients with copd,” European Respiratory Journal, vol. 42, no. 5, pp. 1205–1215, 2013.

[6] J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Activity recognition us- ing cell phone accelerometers,” ACM SigKDD Explorations Newsletter, vol. 12, no. 2, pp. 74–82, 2011.

[7] J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Cell phone-based bio- metric identification,” in Biometrics: Theory Applications and Systems (BTAS), 2010 Fourth IEEE International Conference on. IEEE, 2010, pp. 1–7.

[8] G. M. Weiss and J. W. Lockhart, “Identifying user traits by mining smart phone accelerometer data,” in Proceedings of the Fifth Interna- tional Workshop on Knowledge Discovery from Sensor Data. ACM, 2011, pp. 61–69.

[9] J. W. Lockhart and G. M. Weiss, “Limitations with activity recogni- tion methodology & data sets,” in Proceedings of the 2014 ACM In- ternational Joint Conference on Pervasive and Ubiquitous Computing: Adjunct Publication. ACM, 2014, pp. 747–756.

[10] Q. Cheng, J. Juen, and B. R. Schatz, “Using mobile phones to simu- late pulse oximeters: gait analysis predicts oxygen saturation,” in Pro- ceedings of the 5th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics. ACM, 2014, pp. 331–340.

[11] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg et al., “Scikit-learn: Machine learning in python,” The Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011.

[12] J. Seabold and J. Perktold, “Statsmodels: Econometric and statistical modeling with python,” in Proceedings of the 9th Python in Science Conference, 2010.

[13] N. Cristianini and J. Shawe-Taylor, An introduction to support vector machines and other kernel-based learning methods. Cambridge univer- sity press, 2000.

[14] E. W. Weisstein, “Fast fourier transform,” 2015.

[15] T. K. Ho, “Random decision forests,” in Document Analysis and Recog- nition, 1995., Proceedings of the Third International Conference on, vol. 1. IEEE, 1995, pp. 278–282.

[16] T. Brezmes, J.-L. Gorricho, and J. Cotrina, “Activity recognition from accelerometer data on a mobile phone,” in Distributed computing, arti- ficial intelligence, bioinformatics, soft computing, and ambient assisted living. Springer, 2009, pp. 796–799.

[17] L. Sun, D. Zhang, B. Li, B. Guo, and S. Li, “Activity recognition on an accelerometer embedded mobile phone with varying positions and orientations,” in Ubiquitous intelligence and computing. Springer, 2010, pp. 548–562.

[18] Z. He, Z. Liu, L. Jin, L.-X. Zhen, and J.-C. Huang, “Weightlessness featurea novel feature for single tri-axial accelerometer based activity recognition,” in Pattern Recognition, 2008. ICPR 2008. 19th Interna- tional Conference on. IEEE, 2008, pp. 1–4.

[19] C. Nickel, H. Brandt, and C. Busch, “Benchmarking the performance of svms and hmms for accelerometer-based biometric gait recognition,” in Signal Processing and Information Technology (ISSPIT), 2011 IEEE International Symposium on. IEEE, 2011, pp. 281–286.

[20] N. Ravi, N. Dandekar, P. Mysore, and M. L. Littman, “Activity recog- nition from accelerometer data,” in AAAI, vol. 5, 2005, pp. 1541–1546.

[21] Q. Cheng, J. Juen, Y. Li, V. Prieto-Centurion, J. A. Krishnan, and B. R. Schatz, “Gaittrack: Health monitoring of body motion from spatio- temporal parameters of simple smart phones,” in Proceedings of the In- ternational Conference on Bioinformatics, Computational Biology and Biomedical Informatics. ACM, 2013, p. 897.

[22] A. Brajdic and R. Harle, “Walk detection and step counting on uncon- strained smartphones,” in Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing. ACM, 2013, pp. 225–234.

[23] J. Juen, Q. Cheng, and B. Schatz, “A natural walking monitor for pul- monary patients using mobile phones,” Biomedical and Health Informat- ics, IEEE Journal of, vol. 19, no. 4, pp. 1399–1405, 2015.