Stochastic recognition of human daily activities via hybrid descriptors and random forest using wearable sensors

Abstract

Human daily activity recognition (HDAR) using wearable sensors is an important task for researchers aiming to develop an effective and feasible model which is capable of accurately detecting human motion patterns. These applications provide elderly care, surveillance systems, and wellness tracking. Despite the pervasive use, recognition and monitoring of human physical activities remains inaccurate, which may contribute to negative reactions and feedback. This paper addresses a data-driven approach to recognizing human daily activities in an indoor-outdoor environment. To improve the classification and recognition of human life-log activities (for example, walking, drinking, and exercising), a model is introduced that incorporates pre-processing (such as denoising), hybrid features extraction from four domains, including time, frequency, wavelet, and time-frequency respectively. After that, stochastic gradient descent is exploited to optimize the selected features. The optimal extracted features are advanced to random forest classifiers in order to develop adaptive for human life-log activities. Additionally, the proposed HDAR model is experimentally evaluated on three benchmark datasets, namely, USC-HAD, which is comprised of 12 physical activities, IM-WSHA, which involves 11 life-log activities, and MOTIONSENSE which contains six static and dynamic activities, respectively. The experimental results show that the proposed HDAR method significantly achieves better results and outperforms others in terms of recognition rates of 91.08%, 91.45%, and 93.16% respectively, when the USC-HAD, IM-WSHA, and MOTIONSENE databases are applied.