WiDet

Abstract

To achieve device-free person detection, various types of signal features, such as moving statistics and wavelet representations, have been extracted from the Wi-Fi Received Signal Strength Index (RSSI), whose value fluctuates when human subjects move near the Wi-Fi transceivers. However, these features do not work effectively under different deployments of Wi-Fi transceivers because each transceiver has a unique RSSI fluctuation pattern that depends on its specific wireless channel and hardware characteristics. To address this problem, we present WiDet, a system that uses a deep Convolutional Neural Network (CNN) approach for person detection. The CNN achieves effective and robust detection feature extraction by exploring distinguishable patterns in Wi-Fi RSSI data. With a large number of internal parameters, the CNN can record and recognize the different RSSI fluctuation patterns from different transceivers. We further apply the data augmentation method to improve the algorithm robustness to wireless interferences and pedestrian speed changes. To take advantage of the wide availability of the existing Wi-Fi devices, we design a collaborative sensing technique that can recognize the subject moving directions. To validate the proposed design, we implement a prototype system that consists of three Wi-Fi packet transmitters and one receiver on low-cost off-the-shelf embedded development boards. In a multi-day experiment with a total of 163 walking events, WiDet achieves 94.5% of detection accuracy in detecting pedestrians, which outperforms the moving statistics and the wavelet representation based approaches by 22% and 8%, respectively.