Comparative Analysis of Different Approaches to Human Activity Recognition Based on Accelerometer Signals

Recently, automatic human activity recognition has drawn much attention. On one hand, this is due to the rapid proliferation and cost degradation of a wide variety of sensing hardware. On the other hand there are urgent growing and pressing demands from many application domains such as: in-home health monitoring especially for the elderly, smart cities, safe driving by monitoring and predicting driver’s behavior, healthcare applications, entertainment, assessment of therapy, performance evaluation in sports, etc. In this paper we focus on activities of daily living (ADL), which are routine activities that people tend to do every day without needing assistance. We have used a public dataset of acceleration data collected with a wrist-worn accelerometer for 14 different ADL activities. Our objective is to perform an extensive comparative study of the predictive power of several paradigms to model and classify ADL activities. To the best of our knowledge, almost all techniques for activity recognition are based on methods from the machine learning literature (particularly, supervised learning). Our comparative study widens the scope of techniques that can be used for automatic analysis of human activities and provides a valuation of the relative effectiveness and efficiency of a potentially myriad pool of techniques. We apply two different paradigms for the analysis and classification of daily living activities: (1) techniques based on supervised machine learning and (2) techniques based on estimating the empirical distribution of the time series data and use metric-theoretic techniques to estimate the dissimilarity between two distributions. We used several evaluation metrics including confusion matrices, overall accuracy, sensitivity and specificity for each activity, and relative computational performance. In each approach we used some of the well-known techniques in our experimentation and analysis. For example, in supervised learning we applied both support vector machines and random forests. One of the main conclusions from our analysis is that the simplest techniques, for example, using empirical discrete distributions as models, have almost the best performance in terms of both accuracy and computational efficiency.