Sparse Representation for Device-Free Human Detection and Localization with COTS RFID

Passive human detection and localization is the basis for a broad range of intelligent scenarios including unmanned supermarket, health monitoring, etc. Existing computer vision or wearable sensor based methods though can obtain high precision, they still face some inherent defects, such as privacy issues, battery power limitations. Based on the human movement induced backscattered signal changes, we propose a device-free human detection and localization system on radio-frequency identification (RFID) devices. The system extracts environment-independent features from both RSSI and phase for dynamic monitoring in the first stage, then the target is further located if the moving human is detected. In particular, an overcomplete dictionary is learned when creating the fingerprint library, which helps to make the representation of the location more compact and computationally simple. Moreover, PCA based dimensionality reduction method is then adopted to acquire valid features to determine the final position. Extensive experiments conducted in real-life office and bedroom demonstrate that the proposed system provides high accuracy for human detection and achieves the average distance error of less than 1 m.