Research On Vehicle Tracking Technique In Complex Situations

Tracking for moving vehicles in video surveillance is a challenging problem andplays a key role in intelligent transportation systems (ITS). Digital image processing,combined with advanced scientific and technical achievements in the areas ofcomputer vision, pattern recognition, artificial intelligence and automatic control, hasbeen used to solve a wide range of problems in ITS.For the past few years, much research has been done on vehicle target tracking byscholars at home and abroad. The accuracy of vehicle tracking, however, is still adifficult point due to the variability of vehicles in motion and the complexity ofweather conditions. This paper focuses on these issues and develops a comprehensiveframework to solve these problems.The studies are carried out under the scenario that the vehicles to be tracked areunder complex circumstances. This paper thoroughly investigates the vehiclevariations involving movement at different speeds, partial or even total vehicleblocking, changes in vehicle dimensions, vehicle rotation and deformation, andinterferences of vehicle shaped objects; as well as the complexity of externalenvironment, such as the movement and vibration of monitoring camera, and severeweather conditions, etc. These are inevitable factors in the process of vehicle tracking.A long-lasting and stable tracking method, which remains effective in the precedingsituations, is fundamental to successful vehicle tracking. In this paper, a creative research work is carried out and a robust vehicle trackingalgorithm which is immune to the above-mentioned “complex circumstances” isdeveloped. The details are as below:1. For the consideration of tracking failure caused by fast moving vehicles, thedrawbacks of the mean shift algorithm used for vehicle target tracking are explored.We propose a new mean shift algorithm collaborating with the Kalman filter to trackhigh speed vehicles. The Kalman filter is first applied to determine the vehicleposition, then the well-estimated vehicle position is used for the initial iteration of themean shift algorithm, as a result of which the number of mean shift iterations isreduced and the tracking accuracy is improved. Furthermore, the Kalman gain matrixis calculated and updated repeatedly to improve efficiency while remaining hightracking accuracy. Experimental results show that this algorithm can effectively trackhigh speed moving vehicles with high accuracy. The reduction of time consumptionenables the applicability of real-time vehicle tracking.2. Taking into account the problematic issues involving vehicle partial or eventotal blocking, interference of vehicle shaped objects and tracking difficulties causedby high speed moving, we propose a particle filter tracking method based onsegmentation compensation. When a moving vehicle is blocked or a vehicle shapedobject emerges, the target area is segmented, and each segmentation is trackedindividually. The final vehicle position is determined by adding the vectors of allsegmented regions. The particle compensation algorithm is developed against thetracking problem arising from fast moving vehicles. This method enables the scatteredparticles to cluster together around the true target location. It improves particleutilization efficiency and tracking accuracy, and prevents tracking failure due to highspeed vehicles. Experimental results reveal that the segmentation compensation basedparticle filter algorithm exhibits excellent tracking performance in case vehicles areblocked, interferences of analogues exist and vehicles move at high speed.3. On account of the fact that the process of vehicle edge extraction is susceptibleto noise interferences, which leads to imprecise edge information, we propose an edge extraction algorithm based on Kernel density estimation and mean square error. Thisalgorithm uses two common non-parametric methods so that the detected object edgesand the statistical model of the data distribution are not correlated. Comparisons aremade with the traditional WMW algorithm, KS algorithm and Canny algorithm. Ournew algorithm, applied to images corrupted by different types of noises (Gaussianwhite noises, salt and pepper noises and mixed noises), performs as good as theparametric edge detection algorithms. Even if the actual data distribution isinconsistent with the hypothesised statistical model, this algorithm is still able toprovide satisfactory edge detection performance. Experiments indicate that thisalgorithm possesses strong interference resistant capability in detecting edges of noisyimages.4. In view of the considerable variations of characteristic and dimension oftargeting vehicles during long-term tracking, and the lack of ability to adapt to vehiclerotation and deformation for long-term stable tracking, we develop a new vehicletracking method by combining image edge information with the mean shift algorithm.The vehicle edge information, which has been extracted by the Kernel densityestimation and mean square error based edge detection algorithm, is used to adjust thesize of tracking block adaptively and to enable the real-time updates of vehiclefeatures, then the tracking is accomplished using mean shift algorithm. The proposedalgorithm effectively solves the low tracking accuracy problem caused by the changesof target characteristic resulting from significant variations of vehicle features. It isproved by experiments that this algorithm exhibits outstanding vehicle trackingperformance in “complex situations” such as vehicle shape variation, especially incase of large-scale scaling and rotation, and partly or completely blocked vehicles.