Gesture Recognition with mmWave FMCW Radar Using a Trajectory-Aware Graph Encoder

Gesture recognition using wireless signals has gained significant attention as a promising approach for contactless human-computer interaction. Among various wireless sensing tools, millimeter-wave (mmWave) frequency-modulated continuous wave (FMCW) radar provides rich information such as range, angle, and velocity, making it well-suited for human motion sensing. However, existing studies often underutilize the trajectory information of human body parts, and the transformation of radar data through image processing or resampling often leads to a loss of inherent spatio-temporal patterns. To address these limitations, we propose a graph construction method that effectively preserves spatial and temporal information by representing a point-cloud data of FMCW radar as a directed graph. Based on this representation, we design a lightweight trajectory-aware graph encoder that efficiently processes a directed graph. The proposed graph encoder is readily compatible with various classifiers, making it applicable to broader tasks such as openset recognition which requires detection of unknown classes. Extensive experiments on public datasets demonstrate that our trajectory-aware gesture recognition scheme greatly outperforms existing state-of-the-art schemes under various cross-domain situations.