Wi-Fi CSI-Based Activity Recognition with Adaptive Sampling Rate Selection

Activity recognition methods using Wi-Fi Channel State Information (CSI) have been actively studied in the mobile and ubiquitous computing community. Many prior studies on CSI-based context recognition systems employ CSI data collected at a high and constant sampling rate, resulting in always high computation costs for context recognition. In this study, we propose a CSI-based activity recognition method that adaptively adjusts the sampling rate using reinforcement learning. In the proposed method, the “action” in the reinforcement learning is defined as the selection of a sampling rate of CSI, and the “state” is defined as an intermediate output of a neural network for activity recognition in the environment, which is expected to include information describing the complexity of the current activity. Moreover, we design an activity recognition model that can accept CSI inputs collected at an arbitrary sampling rate in principle, and extract sampling-rate-independent intermediate representations in its intermediate layers, enabling the reinforcement learning agent to switch to an appropriate sampling rate regardless of the current sampling rate. We evaluated the proposed approach using data collected in real environments.