Human Activity Recognition System Based on Channel State Information

Abstract

Human activity recognition (HAR) based on Channel State Information (CSI) has been widely concerned in recent years due to its properties of low cost and privacy protection. Whats more, unlike vision-based and sensor-based HAR systems, it is not sensitive to illumination intensity and does not need any accessory equipment. The detection threshold in existing HAR system is set manually thus require researchers to reset the threshold when environment changes or a new activity needs to be recognized. The utilization of single classifier with poor robustness cannot achieve a high recognition accuracy. We propose a CSI-based Device-free HAR (CDHAR) system to recognize common human activities, such as ‘Walk’, ‘Run’, ‘Sit down’, ‘Squat’ and ‘Fall down’. First of all, CDHAR obtain an adaptive detection threshold to complete the extraction of activity durations. Second, it proposes a random subspace classifier ensemble method for classification. At last, we prototype CDHAR on commodity WiFi devices and evaluate its performance both in typical indoor and outdoor environments. Our results show even experimental site changes, the extraction accuracy rate can achieve 99.80% and 99.60% in outdoor and indoor environment, respectively. Based on the extracted data, the recognition accuracy rate can reach 91.2and 90.2% in outdoor and indoor environment, respectively. CDHAR solves the problem of setting the detection threshold manually and overcomes the low robustness of the single classifier at the same time. Furthermore, the results in above two scenarios demonstrate that CDHAR can also achieve slightly recognition accuracy improvement over existing recognition methods.