Research On Object Detection And Tracking Technology Of Locomotive Workshop Based On Video Surveillance

In recent years,the application of intelligent video surveillance systems has been more and more widely used in enterprises,and has be expanded from basic security fields to workshop production scheduling fields.This thesis comes from the actual demand of CRRC Tangshan Co.,LTD for nine equipments real-time detection and tracking.(train platform,bottom lifting,vacuum sucker,forklift,line work,blue lift,orange lift,yellow lift,mixed color lift).In order to meet business needs,this thesis employs Workshop-YOLOV3(Workshop You Only Look Once)network and improved Deep＿Sort(Deep Simple online realtime tracking)tracking algorithm to realize the object detection and tracking tasks.The main research work is as follows.1.Construct locomotive equipment image dataset,which contains 45 period of videos and 24,496 images.There are nine categories in total:train platform,bottom lifting,vacuum sucker,forklift,line work,blue lift,orange lift,yellow lift,and mixed color lift.Each contains 5 video clips of about 120 s,which has frame rate of 30 FPS(Frames Per Second)and resolution of 1280 × 720.24496 images are extracted from the vedio clips and are labeled with Lableimag software,and a total of 34,326 labeling boxes are obtained.2.Drawing on the idea of dense network and using GIOU(Generalized Intersection Over Union)as the bounding box loss function,Workshop-YOLOV3 network is constructed to achieve object detection of vehicle production equipment in locomotive workshops.The experimental results show that m AP(mean Average Precision)of the proposed network reach 91.32%,7.26% higher than the original model(m AP: 84.06%).In particular,for the case of partial occlusion and near the surveillance edge,19.78% and 31.15% of m AP are improved,respectively.The detection speed of Workshop-yolov3 model can reach 25 FPS,which meets the real-time requirements of video surveillance.3.Improved Deep＿Sort algorithm is designed to implement multi-object tracking,in which object correlation is carried out by using the object displacement strategy and the nearest neighbor object displacement to carry out multi-object matching,and Work-shop YOLOV3 is used as the detector.The experimental results show that the improved multi-object tracking algorithm reach 93.77%,15.35% higher than the original algorithm(78.42%)in the case of tracking after occlusion.The tracking speed of the improved multi-object tracking algorithm can reach 24 FPS,which meets the real-time requirement of video monitoring and tracking.