Design optimization of activity recognition system on an embedded platform

Abstract

Activity Recognition (AR) is a subset of pervasive computing that attempts to identify physical actions performed by a user. Previous sensor-based AR systems involve computation and energy overheads incurred by the use of heterogeneous and large number of sensors, however it is possible to arrive at an optimized system where the design involves optimization of energy consumption through number of sensors, computation through minimal set of features and cost through a nominal hardware platform ideally making it a multidimensional optimization. The above mentioned modelling was reflected in the construction of this optimized system as the design employs a single accelerometer and extracts only 7 time-domain features resulting in ease of computation to classify the activities, thus encouraging it to be inherently deployable on an embedded platform. The system was trained and tested on the accelerometer data acquired from three publicly available datasets. The performance of four chosen machine learning based classification models from an initial set of eight was evaluated, analysed and ranked on the grounds of efficiency and computation. The model was implemented on a Raspberry Pi Zero (USD 5) and the average time for feature computation and the maximum time taken to classify an instance of an activity was found to be 0.015 s and 1.094 s respectively, thus validating the viability of the system on an embedded platform and making it affordable to the population in the low-income groups.