Wearable sensing for activity recognition
Oon Peen Gan · IECON 2017 - 43rd Annual Conference of the IEEE Industrial Electronics Society · 2017
Monitoring workplace activities (what, who, when and where) is beneficial for workforce management. Wearable Recognition enables efficient activities recognition and inferring workplace activities from wearable sensors is attractive but complex. This paper studied the feature extraction of wearable sensory data to uncover the applicability and robustness of wearable sensing approach to activity recognition. It benchmarked a set of time and frequency domain features, namely mean, standard deviation, peak amplitude and energy from the amplitude spectrum extracted from accelerometer in smartphones, for their discriminatory performance in classifying workplace activities. Our experimental results showed that the window size selection reduced variation in time domain feature space but not frequency domain. In our experiments, standard deviation, peak amplitude and energy features outperformed the mean feature by up to 40% in classification accuracy, and using k-nearest neighbor classifier, activity recognition achieved accuracy of about 80%.