Statistical and Time Series Analysis of Accelerometer Signals for Human Activity Recognition

Abstract

Sensor-based human activity recognition HAR has become increasingly more important in our daily lives for a number of reasons. Advances in the sensing capabilities of personal devices have seen unprecedented growth over the past decade. HAR systems have many applications especially in health monitoring, intelligent environments, and smart spaces. Wearable sensors are particularly suited in these areas. This is due to the fact that they have small size, their cost has been steadily decreasing, and they are currently embedded in almost all commodity mobile devices such as smart phones, smart watches, sensory gloves, hand straps, and shoes. In this paper we focus on analyzing sensory accelerometer data collected from wearable devices. And in particular, we study activities of daily living (ADL) which are the activities ordinary people have the ability for doing on a daily basis like eating, moving, individual hygiene, and dressing. To the best of our knowledge most HAR systems are based on supervised machine learning techniques and algorithms, In this paper we widens the scope of techniques that can be used for the automatic analysis of human activities and provide a valuation of the relative effectiveness and efficiency of a potentially myriad pool of techniques. Specifically, we apply two approaches. The first approach is time-aware treating the incoming data in its natural form as a sequential temporal sequence of measurements. The techniques we used are based on time series analysis. The other approach is time-neglectful. It is based on using statistical methods based on goodness-of-fit tests. Our comparative assessment shows that the latter approach has some potential in classification accuracy, though needs further investigation. The time-aware approach gives much better results, though the computational resources required can be prohibitive, so also needs further investigation from that perspective.