Detection and tracking of moving vehicles using affine transform model.

Detection and tracking of moving vehicles plays an important role in transport systems. Based on these functions, it is possible to provide real time routing services to customers who wish to journey between any two nominated points in a traffic network at any time of the day. The emerging intelligent transportation system (ITS) chooses generally to use digital camera for vehicle detection and tracking, instead of analog device (e.g. loop). In order to collect traffic information, such as Degree of Saturation (DOS), traffic incidences etc., video cameras are installed usually to detect or track moving vehicles on a traffic link or intersection. A problem when using a unfixed camera for outdoor environments (for instance, urban traffic) video capture is that the images may be degraded by noises which come both from camera's dithering and intrinsic noise. In some case, when the noises are serious, movable objects even may be overwhelmed completely. For intrinsic noise, most methods proposed treat it as a Gauss by building a statistic model. For camera's dithering, which contribute a global motion to image sequences, is usually compensated by motion estimation. The main point is to estimate the parameters for motion model accurately. In most cases, LS (least square) method is used. But the estimation accuracy is easily spoiled by some singular points with large error. In order to solve this problem, we propose a method based on 2-order Taylor approximation of global warping errors, which refines parameters and so helps to get a more accurate motion model. It is believed that the method discussed in the paper, which is based technically speaking on the computation of a system of 6-order linear equations, is well suited for real-time motion model parameters estimation. The second section presents the motion model, the third section discusses motion estimation, the forth section gives our approach to refine model parameters, the fifth section gives vehicles detection algorithm, and conclusion is made in the last part.