A Reinforcement Learning Based Design of Compressive Sensing Systems for Human Activity Recognition

This paper presents a reinforcement learning based distributed compressive sensing system design method for human activity recognition. This system uses distributed infrared sensors to capture human motion information and aims at representing complex activity scenarios with as little amount of data as possible. Therefore, a set of binary sampling masks are designed to modulate the fields of view (FoV) of sensors and to reduce the amount of measurement data without losing the major features of the target information. The spatial relation between two adjacent sensors is investigated to acquire 3d information with the maximum efficiency. In this work, the main contributions include two parts: (1) design the optimal deployment of distributed sensors and (2) learn the structure of sampling masks by using the policy gradient (PG) based reinforcement learning scheme. Experiment results show that the the proposed system can increase the sensing efficiency and improve the performance of activity recognition.