Toward robust and adaptive pedestrian monitoring using CSI: design, implementation, and evaluation

This work puts the first effort on investigating robust and adaptive pedestrian passing detection and direction recognition based on WiFi Channel State Information (CSI). Specifically, we first give an insight into the challenges as well as opportunities of realizing cross-scenario pedestrian monitoring based on comprehensive analysis of CSI patterns. In light of the findings, we design a novel system framework, which consists of an offline pattern clustering and training module for constructing a unified offline database, and an online adaptive monitoring module for enabling real-time pedestrian passing detection and direction recognition. Further, we propose four mechanisms, including a CSI pre-processing method to enhance system robustness by extracting stable and distinct CSI features, a Two-stage Clustering (TC) method to enable cross-scenario CSI feature classification by segmenting the offline datasets automatically, a unified segmenting and detecting (USD) method to enable adaptive pedestrian passing detection by training a component classifier and a sample classifier, and a dynamic direction calculation (DDC) method to recognize the passing direction based on time of passing estimation, link confidence calculation, and direction indicator calculation. Finally, we implement the system prototype and evaluate the system performance in real-world scenarios. A comprehensive experimental study demonstrates that the proposed framework and the mechanisms can effectively enhance system robustness and adaptiveness on pedestrian monitoring.