Abstract: With the rapid development of autonomous driving technology, how to accurately and quickly recognize traffic command gestures has become an emerging traffic safety technical problem. Kinect is used to extract human skeleton data, and according to the characteristics of traffic gestures, bone angle features and the joint distance features are extracted and compared for analysis. The dynamic time warping (DTW) algorithm is combined with the Kinect for gesture recognition. Aiming at the path singularity problem of DTW algorithm, the shape feature and numerical feature of the sequence are used to improve the DTW algorithm. Meanwhile, considering the influence of unknown actions outside the boundary, an action discrimination method based on the correlation coefficient of matching distance sequence waveforms is proposed. The experimental results show that in terms of traffic command gesture recognition, the recognition accuracy of bone angle features is better than that of joint distance features, and the improved DTW algorithm has higher recognition accuracy than the DTW algorithm and LSTM network, and exhibits better generalization than the DTW algorithm.