Instance-Wise Dynamic Sensor Selection for Human Activity Recognition

Abstract

Human Activity Recognition (HAR) is an important application of smart wearable/mobile systems for many human-centric problems such as healthcare. The multi-sensor synchronous measurement has shown better performance for HAR than a single sensor. However, the multi-sensor setting increases the costs of data transmission, computation and energy. Therefore, the efficient sensor selection to balance recognition accuracy and sensor cost is the critical challenge. In this paper, we propose an Instance-wise Dynamic Sensor Selection (IDSS) method for HAR. Firstly, we formalize this problem as minimizing both activity classification loss and sensor number by dynamically selecting a sparse subset for each instance. Then, IDSS solves the above minimization problem via Markov Decision Process whose policy for sensor selection is learned by exploiting the instance-wise states using Imitation Learning. In order to optimize the parameters of the activity classification model and the sensor selection policy, an algorithm named Mutual DAgger is proposed to alternatively enhance their learning process. To evaluate the performance of IDSS, we conduct experiments on three real-world HAR datasets. The experimental results show that IDSS can effectively reduce the overall sensor number without losing accuracy and outperforms the state-of-the-art methods regarding the combined measurement of accuracy and sensor number.

Optimal sensor channel selection for resource-efficient deep activity recognition

“…Yang et al [29] proposed a sensor channel selection algorithm for shallow learning. The authors addressed the problem by optimizing the performance and number of selected channels via Markov Decision Process.…”

“…Two important points should be clarified about Yang's [29] method. First, a reduction in computational expense only happens if the cost corresponding to the policy network plus the cost involved in the classification of the sparse features does not surpass the cost of classifying the dense features.…”

“…We devise an optimization-based method that presents modularity with respect to the sensor channels -as it is also done in [29]. However, differently, our method does not require zeroing-out inputs corresponding to excluded sensor channels.…”

“…Therefore, in this sense, gains in resource efficiency are more obtainable since sparse tensor algebra is not required. Also differently from [29], memory footprint in our neural network can be reduced when sensor channels are excluded.…”

“…Ours Exhaustive search [1,8,9,30] Sensor channel selection [5,15,17,29] Standard DL methods [10,11,19,20,28] in the shortening of the search space, but it may require domain knowledge.…”