Abstract
<title>Abstract</title> <p>Traffic police command gestures are legally binding visual signals in mixed urban traffic, yet they are difficult for vehicle-mounted perception systems to recognize when rain, snow, or fog degrades image quality and destabilizes skeleton keypoints. This study develops a multi-weather robust spatial-temporal graph convolutional network (MR-STGCN) for recognizing eight traffic police command gestures from two-dimensional skeleton sequences. The network models COCO-format keypoints as a spatial-temporal graph and incorporates two targeted modules. Dynamic regional routing attention (DRRA) evaluates semantic correlations among limb regions and routes attention away from unreliable regions, reducing error propagation caused by occluded or drifting joints. A lightweight multi-scale dilated convolution module (MSDCM) captures both rapid arm swings and longer pose-holding patterns through depthwise separable temporal convolutions. On a self-built multi-weather traffic police command gesture dataset, MR-STGCN achieves 94.2% action-sequence accuracy with 2.85 M parameters, 1221 MB peak memory, and 19.1 ms average action sequence decision time. The results show that regional reliability modeling and multi-scale temporal modeling improve the recognition of similar gestures under adverse weather while maintaining low recognition-module latency and moderate resource consumption.</p>