Research On Human Behavior Recognition Algorithm For Traction Substation Based On 3D Convolution Network And Pose Estimation

High-speed railway is an important infrastructure for the economic and social development of our country,and traction substation is the key power supply facility for the traction power supply system of high-speed railway.In the process of promoting the "unmanned" of the traction substation,its operation safety is more prominent,in which the behavior of the traction substation personnel may endanger the operation safety of the traction substation,this paper aims at the existing method of traction substation human behavior recognition there is a problem of insufficient anti-interference ability and poor robustness,and studies the human behavior recognition algorithm of the traction substation based on 3D convolutional network and attitude estimation,which first screens out the key frames of human behavior through the human body detection model.Then the human posture estimation network is used to estimate the human posture,and the 3D-CNN human behavior recognition network based on optical flow output recognition results,and finally fused by weight adaptive method to output the final human behavior recognition results,which not only takes into account the real-time,but also ensures the recognition accuracy.The main research content of this article is as follows:(1)For the traction substation scene,the TSH dataset and RGB-S dataset are collected and constructed,including different light intensity,angle,and simulated human behavior in the scene.(2)The implementation scheme of human detection in traction substation is studied,and a human detection model based on improved YOLO X-Darknet53 network is designed,which is conducive to extracting the effective characteristics of the target by pruning the channel of the YOLO X-Darknet53 trunk network and integrating the attention CBAM module into the Res Net depth residual network.Experimental results on the COCO 2017 dataset with the original YOLO X-Darknet53 model showed that the average accuracy increased by 4.0% to98.2%.(3)The algorithm of human posture estimation of traction substation is studied,and the human posture estimation model of the lightweight Open Pose-SE network is designed,which is first based on the Open Pose human posture estimation network,and uses the Mobile Net network instead of the VGG-19 standard convolutional neural network,and at the same time integrates the attention SE module,so that the model is lightweight,and the important information features are flexibly emphasized,and the feature channel with low correlation is suppressed.After locating the position of the human body in the image,the characteristics of the key points of the human skeleton are further extracted to obtain accurate human posture information.Experimental results on the COCO 2017 dataset with the original Open Pose model show that the video frame detection rate is increased by 19.0 FPS,and the average accuracy is increased by 4.9%,reaching 89.1%.(4)The human behavior recognition algorithm of traction substation based on 3D convolutional network and attitude estimation is studied.The algorithm outputs the recognition result from the 3D-CNN human behavior recognition network based on optical flow,and fuses the result with the output of the attitude estimation model through the weight adaptation method to output the final human behavior recognition result.The experimental results verify that the recognition accuracy of the human behavior of the proposed algorithm in the traction substation can reach 99.56%,which is 23.36% and 11.36% higher than that of IDT and 3D-CNN algorithms,respectively,which proves the accuracy and effectiveness of the proposed algorithm.