Research On Traffic Police Command Action Recognition Based On Skeleton Sequences

With the rapid development of artificial intelligence,big data,and other technologies,machine vision motion recognition technology has been widely used.Automatic recognition of traffic police command action is an important topic in the field of traffic safety,which helps to enhance the environmental awareness of road auto drive system,and has potential value as a research hotspot.Unlike the recognition of fixed objects such as traffic signs and signals,human movements have the characteristics of flexibility and subjectivity.Based on the analysis of standard traffic police command actions,this paper utilizes key human skeleton nodes to efficiently and concisely represent the dynamic information of human actions,extracts the coordinate information of traffic police command action skeleton nodes,and constructs a convolutional neural network(CNN)model to extract the characteristics of traffic police command action skeleton nodes and perform classification and recognition.To reduce the amount of computation,key frame data processing methods are used to achieve acceptable recognition accuracy using only key frames rather than the entire image sequence while maintaining computational efficiency.The correlation between recognition accuracy and different key frame combinations is analyzed,and it is found that the location of key frames greatly affects recognition accuracy.Among them,three quarters of the video(A3)can provide the highest recognition accuracy,and the average recognition rate at the beginning of the video sequence(A0)is the lowest,which means that the first action of the traffic police command action contains its least obvious features.Compared to previous CNN based recognition algorithms,the accuracy rate of key frames for recognition is close to the CNN recognition effect,but for calculation time,compared to the previous selection of a fixed number of frames,the calculation amount is reduced and the recognition speed is improved.Graph convolution networks can better model structural information.Based on spatiotemporal graph convolution neural networks(ST-GCN),attention mechanisms are introduced to enhance the effective node characteristics of traffic police command actions,starting with the effective improvement of key skeleton node data of traffic police command actions and spatial convolution improvement.The introduction of convolutional attention module(CBAM)further promotes the network’s attention to key skeleton nodes,using graph convolution and traditional 2D convolution to learn potential,higher-level abstract features in the spatial and temporal domains,respectively.Finally,the learned features are mapped to the category label of the current action through a classifier.Experiments show that the convolutional attention module can help improve the efficiency of traffic police command action recognition,and the overall recognition accuracy is higher than that of CNN.