WiAct: A Passive WiFi-Based Human Activity Recognition System

• Published in: IEEE sensors journal Vol. 20; no. 1; pp. 296 - 305
• Main Authors: Yan, Huan, Zhang, Yong, Wang, Yujie, Xu, Kangle
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: activity data cutting; Activity recognition; Adaptive algorithms; Adaptive systems; Artificial neural networks; channel state information; Classification; Correlation; Data mining; Doppler effect; extreme learning machine; Feature extraction; Human activity recognition; Human behavior; Human motion recognition; human-computer interaction; Performance evaluation; Robustness; Wearable technology; Wireless communication; Wireless fidelity
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2019.2938245

Nowadays, human behavior recognition research plays a pivotal role in the field of human-computer interaction. However, comprehensive approaches mainly rely on video camera, ambient sensors or wearable devices, which either require arduous deployment or arouse privacy concerns. In this paper, we propose WiAct, a passive WiFibased human activity recognition system, which explores the correlations between body movement and the amplitude information in Channel State Information (CSI) to classify different activities. The system designs a novel Adaptive Activity Cutting Algorithm (AACA) based on the difference in signal variance between the action and non-action parts, which adjusts the threshold adaptively to achieve the best trade-off between performance and robustness. The Doppler shift correlation value is used as classification features, which is extracted by using the correlation of the WiFi device's antennas. Extreme Learning Machine (ELM) is utilized for activity data classification because of its strong generalization ability and fast learning speed. We implement the WiAct prototype using commercial WiFi equipment and evaluate its performance in real-world environments. In the evaluation, WiAct achieves an average accuracy of 94.2% for distinguishing ten actions. We compare different experimental conditions and classification methods, and the results demonstrate its robustness.