A Wearable Physical Activity Sensor System: Its Classification Algorithm and Performance Comparison of Different Sensor Placements

Abstract

This paper presents a wearable physical activity sensor system and its activity classification algorithm. In addition, we investigate possible combinations of different sensor placements, and identify an optimal combination to achieve the best classification performance. The sensor system consists of several sensor modules that can be synchronized to record the accelerations of diverse motions/activities. In our experiment, three sensor modules are mounted on participants’ hand wrists, waists, and ankles, respectively, to collect seven categories of activity accelerations. The proposed classification algorithm consisting of acceleration acquisition, signal preprocessing, feature generation, and feature reduction, is capable of translating time-series acceleration signals into important time- and frequency-domain feature vectors. The dimension of features is reduced by linear discriminate analysis (LDA), and then the reduced features are sent to a k-nearest neighbor (k-NN) classifier for classification. Our experimental results have successfully validated the effectiveness of the proposed classification algorithm. The best classification accuracy is 96.98% when the sensor modules are placed on hand and ankle simultaneously.