基于FMCW雷达的开集手势识别

Radar-based hand gesture recognition is one of the human-computer interaction technologies that requires the user to perform one of the actions that have been predefined as gestures. The system needs to recognize that the user\'s gesture belongs to the category of the predefined action.In practical scenarios, gesture recognition presents an open set problem, complicated by factors such as incorrect gestures performed by users and environmental noise interference. Consequently, the recognition system must not only accurately classify known gestures but also reject unknown ones. To solve this problem, this paper proposes a gesture recognition system based on FMCW radar that recognizes known gestures, rejects out-of-distribution gestures, and clusters the same gestures (either known or unknown). Within this system, the radar echo containing gesture information undergoes Fourier transformation to generate both a time-frequency map and a range-Doppler map. These maps are then input into a two-channel network, which is trained using cross-domain dataset training and metric learning techniques. The system achieves an accuracy rate of 99% for six known gestures and a rejection rate of 95% for unknown gestures.