Traffic Control Gesture Recognition Based on Masked Decoupling Adaptive Graph Convolution Network
Xin Xu, Yuelei Xu, Zhaoxiang Zhang, Kun Gong, Liheng Dong, Linhua Ma · 2023
To promote the application of visual pattern recognition technology in the field of intelligent transportation systems. In this paper, we proposed a novel method, i.e., masked decoupling adaptive graph convolution network (MDe-GCN), which can be utilized to classify traffic control gesture signals. Based on the adaptive graph convolution network, the decoupling graph generation method was adopted to obtain channel-by-channel adaptive topology and introduce spatial/temporal attention mechanisms. Besides, this study innovatively proposed a regularized training method for adaptive graph convolutional networks by generating random graph masks, which can cut off the information transfer between some joint pairs during training. In the experiment, a Chinese traffic control gesture dataset (CTCGD) was collected from filming three actors performing eight traffic control gestures. Through the test, MDe-GCN achieved an action classification performance exceeds other GCNs, with an accuracy of 90.71% in CTCGD.