Study On The Vehicle Target Trajectory Extraction Algorithm

The boom of road traffic industry promotes the development of intelligent transportationtechnology, and the target trajectory extraction is one of important parts of the intelligenttraffic technology. As the macroscopic characteristic of the target motion, vehicle targettrajectory contains rich vehicle motive information, and is the basis of the target vehiclebehavior analysis, the traffic parameters detection and traffic condition detection. What’smore, it can influence the accuracy rate of that detection in a certain sense. The premise toaccurately extract and describe target motion trajectory is that the target must bedistinguished from the background sequence and stably tracked.This paper focus on the study of vehicle target motion trajectory extraction algorithm. Toimprove the accuracy and stability of vehicle tracking and obtain the satisfied target motiontrajectory in the various situations such as scene variation, illumination variation, vehicleocclusion and scale variation, tracking algorithms based on target segmentation, based onlocal characteristics, and especially based on feature points are studied in detail in this thesis.The paper analyses current moving target detection algorithms, using twice results ofthe frame difference gotten by one image and another one before three frame to detectsegment moving target, and then using SAD criterion to achieve target tracking. It tests thetrack algorithm under a variety of situations. It also reviews the common characteristics oftarget, introduces the classical SIFT and SURF descriptors, and uses the SURF algorithm totrack targets.Using improved frame difference method for target detection, to make feature cornersselecting areas. Moravec algorithm is further improved to extract better robustness corner. Akind of dynamic corner selection algorithm is designed, giving priority to higher qualitycorners, improving the phenomenon of too much corners and dense distribution and avoidingsaving excessive number of corners on the same target. In terms of template design, thesmooth areas of the basic template are discarded; using and the non-smooth areas build a newmatch template. And according to the vehicle deformation, the template size is adjusted. Anew template data update algorithm is proposed, taking use of the distance between the matchpoints. It solves the problem of error accumulation and matching point drifting causing by the basic template data update algorithm and improves the accuracy of tracking. Using Kalmanalgorithm for target motion estimation makes tracking trajectories smoother. Utilizing thepredicting location of matching points, can narrow the search range of the target and reducethe amount of calculation.In a variety of scenarios to test tracking algorithm based on feature points, the resultsshow that the tracking algorithm based on feature points proposed in the paper canaccurately track target and extract target trajectory in real-time.