Research On The Technology Of Detection And Tracking On Vehicles For Violation Of Traffic Line Based On Vision

Technology of detection and tracking on vehicles for violation of traffic line is is the critical part of the ITS-Intelligent Transportation System.It has a very important significance in regularizing driver's behavior,reducing traffic accident rate,improving the efficiency of traffic,etc.This paper studies the technology of technology of detection and tracking on vehicles for violation of traffic line based on vision.Firstly,for the negtive effect on vehicle detection from car's short period stationary state at the place of intersections and traffic lights,this paper proposes a background update method based on partition and neighborhood information.First,divide the image into two kinds of partitions: partitions with target and partitions without target,then use average background method to update the background of partitions without target,and adjust brightness of partitions with target according to neighborhood information,finally detect moving target throung the difference of color and brightness between current image and background.According to the experiment,compared with average background method,the foreground extraction rate of the algorithm proposed in this paper increased by 26.9%,the rate compared with GMM mothod increased by 27.5% and conpared with ViBe method increased by 31.1%.For the limitations of normalized cross-correlation method,this paper proposed a shadow detection method based on the integration of normalized cross-correlation method and symmetric cross-entropy,and then reduce detection error rate through neighborhood information of pixels on the contour of the shadows.Copmpared with normalized cross-correlation method and SVM method,the foreground retention of the algorithm proposed in this paper is improved by 40.4% and 6.3%.This paper solves the problem of the negtive effect on vehicle detection from car's short period stationary state,and improves the moving target detection,which has a better foreground detection result and can satisfy the real-time constraints.Secondly,for the problem that conventional algorithm of vehicles detection for violation of traffic line cannot apply to violation detection in diversion line area with frequent traffic jam,a detection algorithm on vehicles for violation of traffic line in diversion line area based on superpixel segmentation and shadow detection is proposed through the region features of the car bonnet and shadow underneath the car.First,achieve the coarse position box througn the moving foreground of the diversion line area and simple calibration,then extract car bonnet area through superpixel segmentation to get exact size of the car in horizontal direction.Further more,achieve exact size of the car in vertical direction according to shadow detection and finally get precise position box of the target.According to the experiment,the accuracy of the algorithm for violation of traffic line in diversion line area,which proposed in this paper,is more than 90%.In the end,aming at the problem of tracking failure and error caused by non-uniform distribution of characteristic points,according to the relative position constraint of characteristic points and tarcking box,this paper proposes a tracking box update method based on corresponding relation of the characteristic points between frames,which integrats the corresponding relation of the characteristic points between frames,the scaling transformation relation of tarcking box between frams,and the position of the tracking box center in the last frame to get the current position of the tracking box center.In addition,improve real-time performance and reduce computation through Kalman filter.Further more,according to the distribute regulation of moving foreground in car area,this paper proposes a screening method of characteristic points based on occlusion judgment.According to the experiment,compared with TLD method,KCF method and particle filter method,the tracking accuracy of the proposed algorithm is improved by 19.7%,14.5% and 33.6%.The proposed algorithm alse solve the problem of occlusion and can satisfy the real-time constraints.