Research And System Implementation Of Vehicle Tracking Algorithm For Less Amount Of Labeled Information

Object tracking mainly segments the target image and background image on the basis of target detection,and then locates and recognizes the target.This process according to the target position,image pixels,features and other information in the adjacent image frame,so as to judge the target sequence number in the video and analyze the movement track of the target.Target tracking algorithm mainly includes target characteristics,tracking model and target search strategy,and the tracking model is the core to determine the effect of tracking algorithm.However,at this stage,most of the target tracking tasks are aimed at pedestrians with enough labeled information in the dataset,while the research on rigid objects such as vehicles with less labeled information is relatively weak.The topic of this thesis is the research and system implementation of vehicle tracking algorithm for less amount of labeled information.The research focuses on vehicle target tracking.Two improvement ideas are proposed for the problem of insufficient tracking accuracy of vehicle targets in the existing road traffic monitoring vehicle information analysis platform in the video scene and the lack of correlation between tracking and attribute information leading to false detection:use attention mechanism to pay attention to vehicle specific details,improve the accuracy of vehicle target tracking in the case of insufficient labeled information;And bind and verify the vehicle target track and attribute information to realize information association,and improve the reliability of vehicle attribute detection on the platform.The topic of this research comes from the project of BUPT-China Telecom Visual Intelligence Joint Laboratory:research and development of vehicle analysis algorithm,and improvement of the "intelligent traffic analysis platform" of the project.The main research contents and innovative achievements of this research are as follows:1.Aiming at the problem that the vehicle target in the road traffic surveillance video lacks rich standard information and is greatly affected by the change of illumination and scale on the appearance,this research proposes a vehicle tracking algorithm for less amount of labeled information,and designs a single-step tracking algorithm network based on the variable DETR(DEtection TRansformer)algorithm.In the target detection stage,attention mechanism can be used to pay more attention to the specific details of the vehicle(such as the outer frame,the edge of the vehicle lamp,etc.),It makes up for the defects in feature matching caused by the lack of labels,and reduces the difficulty of detecting the impact of illumination,scale and complex motion problems on vehicle appearance.For each target vehicle,focus on the more meaningful spatial location and the location containing local information.The design of the tracker preserves the frame information and avoids occlusion and loss;Support high frame rate tracking,give full play to the advantages of fast processing speed of single-step tracking algorithm,improve the accuracy without affecting the speed,and finally achieve the tracking accuracy of 59.9%MOTA(Multiple Object Tracking Accuracy)and 73.2%IDF1(ID F1score).2.For the problem that the existing road traffic monitoring vehicle information analysis platform lacks of vehicle-tracking,and relies too much on the target detection results,and is affected by the lack of vehicle labeled information and the lack of verification of the one-time output vehicle structural information.This research implements a vehicle tracking system for less amount of labeled information based on the tracking results of the single-step tracking algorithm based on the variable DETR algorithm,combining vehicle tracking and attribute detection,The object ID and the license plate detection result with the highest frequency on the track,that is,the license plate with the highest confidence,are jointly matched as verification to realize the association and reuse of information,maximize the reliability of the vehicle’s structured information output in the video,reduce the false detection of ID and license plate,and do not add additional higher requirements for labeling information.The system achieves an average accuracy of 74.94%AP(Average Precision)for multi-scale vehicle target detection,and achieves 94.55%system output accuracy for vehicle structured information based on tracking.It gives full play to the advantages of the consistency and integrity of the traffic monitoring video,effectively reduces the massive redundancy of the background error detection data,and intuitively,efficiently and intelligently manages the vehicle structured information in various traffic scenes based on vehicle tracking and attribute detection.Finally,the "intelligent traffic analysis platform" has been improved to support the four functions of vehicle appearance information identification based on tracking,vehicle behavior analysis,vehicle data statistics and vehicle comprehensive information query,broke through the limitation of picture search and support video car search.