Wi-Tar: Object Detection System Based on CSI Ratio

In recent years, public safety has gradually become the focus of society. Generally, detecting the object requires specialized and expensive equipment, which is not easy to be widely deployed. WiFi is generally used in wireless sensing as a low-cost, convenient, and harmless technology. Existing WiFi-based object detection systems suffer from multipath effects and hardware defects, resulting in a significant amount of noise in the Channel State Information (CSI) extracted from the physical layer. In this paper, we propose Wi-Tar, a novel system that utilizes the quotient of CSI value as the base signal. Wi-Tar analyzes their differences for different materials by the reconstructed CSI complex. For objects of the same material, we design a feature to eliminate the interference of object size and develop a Multi-Channel Integrated Convolutional Neural Networks and Long Short-Term Memory (MCICNN-LSTM) structure. MCICNN-LSTM fully leverages the characteristics of Multiple-input Multiple-output (MIMO) technology, achieving high accuracy. We implement this system in commercial WiFi devices and evaluate its performance in 3 indoor scenarios. Extensive experimental results demonstrate that the recognition accuracy of three materials is over 99.5% in a 2-meter detecting distance, and the average recognition accuracy of objects of the same material exceeds 99%. Even when these objects are placed in different occlusions, Wi-Tar can still accurately distinguish them.