Micro Gesture Recognition with Multi-Dimensional Feature Fusion and CQ-MobileNetV3 Using FMCW Radar

Radar-based gesture recognition technology has gained increasing attention in the context of contactless human-computer interaction (HCI). Micro gestures have smaller motion amplitudes and shorter duration compared with traditional gestures, which increases the difficulty of motion feature extraction. In addition, improving recognition accuracy while maintaining low computational and storage costs is also a challenge. In this paper, a micro gesture recognition method combining multi-dimensional feature fusion and a lightweight CQ-MobileNetV3 network is proposed. For feature extraction, the range-time map, velocity-time map, and angle-time map of gestures are first constructed. Then, normalization and adaptive filtering are performed to refine the three maps. Finally, the three refined maps are fused to form a range-velocity-angle-time map, which can accurately describe the motion characteristics of gestures. For recognition, a lightweight CQ-MobileNetV3 network is designed. First, the network structure of MobileNetV3 is optimized to reduce computational complexity. Then, the improved convolutional block attention module (CBAM) and the improved self-attention (SA) module are constructed and integrated into different bottleneck blocks to improve recognition accuracy. A series of experiments are conducted with a 77 GHz frequency-modulated continuous wave (FMCW) radar. The results indicate that CQ-MobileNetV3 achieves a recognition accuracy of 97.16% for 14 micro gestures, with a parameter count of 0.207 M and a computational complexity of 0.027 GFLOPs, surpassing several other deep neural networks.