Research On Non-motor Vehicle Detection And Tracking Technology For Traffic Video

There are a large number of non-motor vehicles in urban road traffic,and it is easy to cause traffic accidents caused by illegal driving behavior,which poses a threat to the safety of oneself and others.In recent years,cities have invested a lot of human and material resources to strengthen the management of non-motor vehicle violations.With the rapid development of artificial intelligence technology,traffic governance model tends to be unmanned,digital and intelligent.Aiming at the intelligent supervision of non-motor vehicle violations,this paper proposes a non-motor vehicle detection algorithm(YOLOv5_VDE),which integrates VovNet network and deform convolution,and combines ByteTrack multi-object tracking model to design a traffic intelligent identification system for detecting non-motor vehicle violations.The main research content of this paper is as follows:(1)Improved improved object detection algorithms’ ability to detect non-motor vehicles.Aiming at the situation that non-motor vehicles occupy relatively few pixels in the traffic screen,different degrees of mutual occlusion between vehicles and insufficient features caused by driver’s self-occlusion behavior,a new backbone network(CSPVovNet network)is proposed for feature extraction based on the characteristics of VovNet network and the original network,so as to improve the reuse of low-dimensional features by the network and enhance the feature expression ability of the feature extraction network.Deformable convolution is introduced to solve the non-rigid transformation of non-motor vehicles at different shooting angles and heights of traffic video,and improve the robustness of the network model to the non-rigid transformation of target objects.Finally,the EIoU_loss is used as the training convergence speed of the regression localization loss improvement model,and the final model detection accuracy is improved by 4.14%compared with the original network.(2)Optimize and enhance the tracking effect of multi-obj ective tracking algorithm on nonmotor vehicles.In order to alleviate the problem of target ID transformation and improve the tracking stability,the improved detection model is used to replace the original detector in the ByteTrack multi-target tracking algorithm model.The multi-object tracking algorithm before and after the improvement is compared with the open source pedestrian tracking dataset MOT20 and the self-made non-motor vehicle tracking dataset NVT.The experiment shows that the multiobject tracking model after the improvement of the detector has obvious improvement on the frequency of ID transformation and the tracking effect under the condition of short occlusion.(3)Design an intelligent non-motor vehicle violation recognition system.The design of the system is divided into two parts:the front end of the system is responsible for the input of video data,the setting of the road environment and the display of output results;The system backend uses the improved object detection and tracking algorithm to process video data.Set rules for determining non-motor vehicle violations to identify non-motor vehicle wrong-way driving and occupancy of motor lanes in real traffic scenarios.(4)Design intelligent non-motor vehicle violation recognition system.The design of the system is divided into two parts,the front and back end,which are responsible for the input of video data,the setting of road environment,the display of output results and the use of improved target detection and tracking algorithm for video data processing.Combined with the set rules of non-motor vehicle violation,the intelligent identification of non-motor vehicle reverse driving and occupation of motor lane in real traffic scenes is finally realized.