Energy efficient human activity recognition using wearable sensors

Computational and power efficiency is one of the crucial enabling factors to wearable device based human activity recognition (HAR) system. However, limited research efforts in literature have been available toward reducing theses costs without loss of accuracy. In this paper, we propose an improved random forest (RF) based HAR system for elderly-care. The system extracts three kinds of pairwise correlation features in hybrid sliding windows, and uses location information to enhance the recognition performance. A mutual information based feature selection is adopted to optimize the recognition of confused local set of activities. A new random feature selection strategy for each node in RF enables the proposed system to reduce the number of trees while maintaining the recognition accuracy. Numerical experiments show that the proposed method can predict 10 types of activities with 93.01% accuracy and 74.9% reduction of energy consumption. Furthermore, the fall detection accuracy in this proposed system can reach up to 99%.