Adaptive Compressed Classification for Energy Efficient Activity Recognition in Wireless Body Sensor Networks

Energy efficiency for activity recognition is a challenging task in the long-term health monitoring applications. To save the energy in wireless body sensor networks, the adaptive compressed classification framework for activity recognition is proposed. The proposed mechanism estimates the minimum activity-specific compression ratio subject to given recognition accuracy and adaptively adjusts the number of samples taken by sensor nodes based on the feedback from the result of activity recognition. A brute force algorithm is utilized to quickly find the optimal trade-off the power consumption and classification performance. With experiments on real-world activity datasets, we demonstrate 90.5% recognition accuracy with energy consumption reduced 28%.