Research On Target Detection And Tracking Algorithms Based On Traffic Video

With the development of smart cities and intelligent transportation systems,effective target detection and tracking technology has attracted the attention of scholars at home and abroad.Especially in intelligent traffic surveillance,the similarity of vehicle characteristics,the occlusion,illumination,rotation,scale,attitude change and external environment change during vehicle driving bring great challenges to target detection and tracking technology.How to obtain effective information and analyze traffic status from massive traffic videos,and finally realize intelligent control and intelligent management are the technical problems faced by intelligent transportation systems.This paper studies the algorithm of target detection and tracking in traffic video.The main research contents are as follows:(1)The principle and implementation process of several traditional target detection and tracking algorithms are studied.The advantages and limitations of the traditional target detection and tracking algorithms are analyzed,and the performance of different algorithms is compared.Secondly,the TLD(Tracking-Learning-Detection)target tracking algorithm is studied.The TLD algorithm combines the traditional detection technology with the tracking technology to solve the slight deformation,illumination and partial occlusion problems of the target during the long-term tracking process.An improved online learning mechanism has been added to make the overall algorithm more stable and reliable.(2)In view of the large-scale rotation,severe occlusion,and sudden change of illumination in the face of the TLD algorithm,the accuracy of tracking is difficult to guarantee.Two improvements are made to the tracking module:on the one hand,LK(Lucas-Kanade)optical flow method is more sensitive to illumination,and the generated tracking points are not feature points,and can not be used to represent the target,so it is easy to track in the process of tracking.To solve the problem of tracking drift,the ORB(Oriented FAST and Rotated BRIEF)feature detection algorithm and ROSAC optimization algorithm are combined to track the target on the basis of the original tracking module.On the other hand,aiming at the problem that TLD algorithm is easy to fail in tracking whenthe target is occluded or disappeared in a large area,a Kalman filter is proposed to predict the target position when TLD tracking fails.(3)Aiming at the problem that the number of scanning sub-windows of TLD algorithm detection module affects the real-time performance of the system,an improved algorithm of Vibe(Visual Background Extractor)foreground detection with adaptive updating of background model is proposed before the detection module,which replaces global scanning with local scanning.Finally,the experimental results and analysis verify the reliability and effectiveness of the improved algorithm.