| With the accelerated development of deep learning technology in the area of intelligent driving,the demand for detection and recognition technology of traffic police gesture has presented a great research significance.In the task for recognition of traffic police gesture,the rapidity of extracting the skeletal features of traffic police and the accuracy of recognizing continuous movements are the key issues to be overcome in the research.In this paper,we address the practical needs of dynamic gesture recognition for traffic police officers,and carry out improvement and research on the algorithm for pose estimation and gesture recognition of traffic police based on deep learning.The main work is as follows:This paper proposes a method to improve the OpenPose network for pose estimation of traffic police,which is based on the problems of difficult feature extraction and slow detection speed of the traffic police pose estimation algorithm.Using Mobile Net as a backbone network for signature acquisition of traffic police pose.which optimizes the performance degradation of the model with the deepening of the hierarchy,reduces the number of parameters of the network and accelerates the computation of features inside the backbone network;the model is improved from a parallel structure to a series-parallel peer structure through a jump connection mechanism,which realized the sharing of inner parameters of the organize,decreased the complexity of the model,and improved the realtime performance of the skeleton detection of traffic police.Through experimental validation,the improved model demonstrates good detection results and provides a practical method with high robustness and real-time performance for the traffic police pose estimation problem.In arrange to illuminate the issues of low accuracy and destitute robustness for recognition of traffic police gesture,this paper designs an ST-GCN-based method for dynamic gesture recognition of traffic police.Using ST-GCN as the main network,the skeletal topological graph structure of traffic police is optimised,and skeletal features are extracted from two dimensions: space and time,respectively;fusing spatial attention mechanism,the skeletal connectivity structure of traffic police is optimised by updating the graph attention matrix to highlight the effective spatial features in dynamic gestures;adding temporal attention mechanism,giving different degrees of attention to channel information,and enhancing the learning of core action in gesture features.The results show that the STGCN gesture recognition network with the fused attention mechanism achieves higher recognition results,with an average recognition precision of 88.82%,which is 5.79% higher than that of the classic ST-GCN.It is verified that the ST-GCN algorithm integrated with attention mechanism has a good performance in recognition of traffic police gesture.According to the task requirements for detection and recognition of traffic police gesture algorithms in real traffic,a traffic environment-oriented traffic police gesture recognition system is designed.The YOLOv5 s network is used to detect the area of the traffic policeman in view of the interference in the environment,and its position is tracked by the Kalman filtering algorithm to improve the effectiveness of the system’s application in the traffic environment;a graphical interactive interface is built to improve the readability and aesthetics of the recognition results.The system has been tested in both video and online modes,demonstrating its excellent performance and providing a feasible solution for the study within the intelligent driving field for recognition technology of traffic police gesture. |
