Recurrent Graph Convolutional Network for Skeleton-Based Abnormal Driving Behavior Recognition

Abnormal driving behavior recognition is important in driving and traffic safety. Currently, skeleton-based action recognition has achieved significant improvement. However, how to effectively recognize abnormal driving behavior is still challenging in real applications, especially for subtle and similar behaviors. In this work, we propose a novel recurrent graph convolution network, which combines spatiotemporal graph convolutional networks and recurrent neural networks. First, we design a new spatial topological graph that includes the joints of the hands and face, which is advantageous to recognize subtle abnormal driving behaviors, such as yawning. Second, the proposed network can extract discriminative spatial and temporal representation features of the segmented skeleton sequences. Our method achieves an accuracy of 90.04% on the dataset collected by ourselves. Moreover, experiments on the Kinetics dataset verify the generalization ability of our method.