Research And Implementation Of Online And Low-Latency Activity Signal Segmentation For WiFi Sensing

WiFi sensing utilizes the impact of target motion on the propagation of ubiquitous WiFi signals to infer target behavior states,and uses the change in Channel State Information(CSI)to capture the occurrence of target motion.It mainly includes two stages:activity signal segmentation and activity classification.The premise of accurate activity classification is to accurately extract the CSI segments corresponding to the target activity from the CSI sequence.Currently,activity signal segmentation can be achieved through threshold-based methods and deep-learning based methods.Threshold-based methods judge the start and end of an activity through CSI changes that exceed a certain threshold,but the selection of the threshold is highly empirical and requires adjustment based on activity granularity and environment.Activities with different granularity may require different or even conflicting thresholds.Due to the inability to perceive granularity,using preset thresholds may fail.Deep-learning based methods can perceive activity granularity through deep-learning models,but due to high overhead,it is difficult to meet the low latency requirements of online activity signal segmentation in real-world scenarios.This paper proposes an activity granularity-aware metric based on moving variance and an online low latency activity signal segmentation method based on threshold.Through the granularity-aware metric,this paper solves the problem of existing threshold-based methods not being able to perceive granularity and the high cost of deep-learning based methods,and achieves activity granularity feature extraction before signal segmentation.The segmentation method proposed in this paper achieves fast and accurate online activity extraction.The core segmentation algorithm automatically determines the optimal threshold according to the activity granularity-aware metric and completes the segmentation,eliminating the empirical dependence on threshold setting,and solving the problem of unknown threshold calculation range for online activity signals.To avoid the impact of link signal quality on segmentation performance,this paper proposes a link signal quality score and a heuristic link selection mechanism,which automatically selects links as needed based on the result of the core segmentation algorithm.This paper significantly improves the efficiency of existing activity signal segmentation methods and achieves online low latency active signal segmentation.Experimental results show that compared with the existing threshold-based method,this method improves segmentation accuracy by up to 15%,significantly reducing the dependence on threshold setting experience.Compared with the deep learning-based method,this method achieves similar segmentation accuracy,but reduces data processing time by 97%,significantly reduces latency and dependence on hardware resources.