Moving Vehicle Tracking Based On Computer Vision

A new method of moving object detection and tracking from dynamical image sequence is proposed in this paper. This method uses Otsu to achieve the adaptive selection of high threshold in Canny operator. The modified Canny operator is used to detect the object contour from difference images. And this detected contour is then used as the template of partial Hausdorff distance matching. This method predicts the position of moving object in next frame by grey prediction model GM(1,1), so it can enhance matching efficiency by taking such position as reference point to locate the search region.The results of tracking depend on motion feature model and prediction mechanism. Simplification of the motion feature model in this paper relies on the extraction of moving vehicle by Otsu-based Canny operator. If there are more pseudo-edges under the moving scene, the motion feature model extracted directly by Canny operator is not simplifier than model extracted by Otsu-based Canny operator. So in this paper we use Otsu-based Canny operator to extract the contour of moving vehicle on the basis of motion information, which can obviously minish the calculation scale of Hausdorff distance between point sets and shorten the matching time of template in image.As for dynamical image sequence, vision tracking of moving vehicle is realized through continuously applying matching algorithm to each frame in image sequence. But just matching algorithm is not enough, so the prediction model must be added while tracking. The prediction model can estimate and predict motion state of vehicle in current frame according to historical information of vehicle motion. The prediction mechanism improves not only the speed of matching and locating, but also the robust performance of locating.It can be seen from the experimental results that grey prediction model GM(1,1), even without the pre-supposition of motion states, has better results thanα-β-γfilter. Also it can be found that forα-β-γfilter the prediction error increases rapidly with time, but for grey model it is small and stable. It can be known from the experimental results that as for some nonlinear motion trajectories such as curving channel where the vehicle is located, the grey model can reduce the influence on prediction of the whole motion trend imposed by randomicity of sampled data. So under limitation of information grey system model can maintain minor error stably. Grey model is closer to real motion trajectory, and better reflects the motion trend of object. Obviously grey model is better than such linear model asα-β-γfilter.