Study On A New Vehicle Tracking Method Based On SIFT Features Matchment Of Corners

Intelligent Transportation Systems (ITS) is one of modern transportationdevelopment direction, detecting and tracking of the moving vehicles is the basicissue in Intelligent Transportation Systems. On the basis of the analysis andsummary of current existing vehicle detection and tracking technology, this paperaimed at a new vehicle tracking method based on SIFT features matchment ofcorners in static background. Main study work of this paper can be summarized asfollows:(1) Detection and location of the corners. On the basis of the existing vehicledetection method, Harris corner detection algorithm is used to extract the cornerpoints-important local features of the image. Since corner points are of rotationinvariance, so detection of such corner points were almost not subject to theinfluence of the light conditions.128-dimensional SIFT feature vectors were thencalculated to describe the corners. SIFT features remain unchanged on zoom, rotate,scale and brightness changes, therefore, the proposed algorithm can adapt tovariety of lighting conditions. Because the use of the feature point matching, it canobtain a good tracking results when the vehicles are partial occluded.(2) Formation of vehicle corners trajectory. Euclidean distance of the128-dimensional SIFT feature vectors was calculated to indicate the similarity ofcorner points in two consecutive images. Dual-way matching method was used toobtain the matching relation of corner points for the formation of corner trajectories.In order to meet the requirements of real-time operation, while tracking, Kalmanfilter was used to estimate the motion state of trajectories to narrowing the scope ofmatchment. To improve the tracking stability, the estimated point was used as theposition of the trajectory when trajectory can’t be matched in the present frame. Withthe purpose to get the speed of the track and calculate the size of vehicles,homography based camera calibration method was adopted to convert the imagecoordinate system and the world coordinate system. (3)Research on corner trajectories grouping method for vehicle data. Thedifference of trajectory position, speed, direction and moving distance weremeasured to indicate the similarity degree of different trajectories, trajectories werethen classified and grouped by similarity degrees. Since vehicle movementinformation was fully utilized, the proposed algorithm is capable of distinguishingoccluded vehicles.This proposed method has good adaptability of ambient illumination, and canalso distinguish and track the partially occluded vehicles at a certain degree. Smallsample experimental results showed that the proposed vehicle tracking methodachieved a detection rate of98.09%, the undetected rate of2.0%and false detectionrate of1.85%for different car density.