Research On Human Body Recognition And Tracking Algorithm For Railway Patrol Robot

As an important infrastructure of the country,railways are the backbone of the comprehensive transportation system and an effective transportation method to improve resource and energy utilization efficiency.The safety of railways is crucial.At present,the safety of railways mainly relies on manual patrol,but this method has certain limitations.With the popularization of intelligent ideas,the issue of using patrol robots combined with intelligent algorithms to replace manual Manned patrol has attracted much attention.Among them,the detection and tracking algorithms based on deep learning in intelligent algorithms have high accuracy,but there are problems such as excessive model computation and poor real-time performance.This article focuses on the usage scenarios of patrol robots and conducts research on the combination of deep learning based object detection and tracking algorithms with practical application scenarios.The main work is as follows:A target recognition algorithm based on YOLOv3 model improvement is proposed to address the high computational complexity,poor real-time performance,and high missed detection rate of current target recognition algorithms.Drawing inspiration from the idea of convolutional separable operations(DW)in Ghost Net,the Ghost Module is used instead of Conv2 d to divide multi-channel data into multiple groups for simultaneous processing,achieving the goal of reducing network parameters;Add a feature output layer on top of the three feature output layers in the original network to reduce the missed detection rate of target recognition;Simultaneously using the k-means algorithm to calculate the new prior box size;Upgrade the original IOU calculation method to DIOU calculation,add penalty terms,and accelerate the convergence of the model.The experimental results show that in the VOC dataset,compared with the original algorithm,the improved algorithm reduces the m AP value from83.52% to 79.85%,a decrease of 3.67%,while ensuring model complexity;Reduced the network parameters by half,improved the detection speed of the model,and increased the video detection frame from 10 fps to 17 fps,reducing the missed detection rate.In response to the requirement for real-time tracking algorithms,GOTURN is used for target tracking in this project.The target position of the current frame is predicted based on the target characteristics of the previous frame,and Laplace is used to extract the features of the boxed target to learn the general relationship between appearance and motion for offline learning.Only a feedforward of the network is needed to directly regression to the position of the target object.Run at a speed of 100 FPS on the GPU of Open CV.However,the GOTURN algorithm has the problem of being unable to autonomously continue tracking after target tracking is lost.The recognition results of YOLOv3-G are fed back to the GOTURN network to solve this problem.Through experiments,it has been found that as long as the target object is not completely occluded and does not move too fast,the target will be continuously tracked;Using the YOLOv3-G recognition box for feedback,if the target is lost and reappears,the network will rebuild the model for tracking.This article applies the improved YOLOv3 object detection algorithm and GOTURN object tracking algorithm to the patrol robot system,and conducts experiments in outdoor venues,mainly modifying the backbone network,FPN part,and IOU calculation method of YOLOv3;Feed back the recognition results of YOLOv3-G to the GOTURN target tracking algorithm,achieving the goal of automatically recovering tracking after target loss.By verifying that the improved detection algorithm can meet the detection requirements in the case of target occlusion,the tracking algorithm used can meet real-time requirements.