The Research Of Moving Vehicles Tracking Technology Based On Machine Vision

Vision-based target tracking technology is a significant application of machine vision. It has attracted much attention in the research of Intelligent Transportation System(ITS). It is also one of the important research topics in the application of intelligent transportation monitoring. By tracking the moving vehicles we can extract the parameters of the traffic and detect the traffic issues automatically. During the vehicle tracking period, mutual covers of the vehicles, alteration of the light, influence of the complex background as well as the randomness of vehicle movement often happen, which affect the vehicle detection precision directly. How to track the vehicles steadily with complex background, mutual covers of the vehicles, alteration of the light and randomness of vehicle movement is the key aspect of this thesis.In terms of the maneuverability of the vehicles, we build a "current" statistic model for moving targets. It has been widely accepted that the "current" statistic model performs well when tracking moving target but for targets with constant speed or targets with low speed, its tracking ability is poor. In order to solve this problem, this thesis presents an improved "current" statistic model. By adjusting the maximal acceleration, we overcome the drawback of slow convergence or divergence caused by constant acceleration and improve the tracking accuracy.Traditional Kalman Filter is not suitable for moving target whose model is inaccurate. And it will even cause failure in target tracking. Take this phenomenon into account, in this thesis we combine adaptive Kalman Filter with "current" statistic model to build the state filter algorithm. Using the relation between acceleration and predictive state, the maximal acceleration as well as the variance of the maximal acceleration will adapt with the situation of the moving target. The variance of the system noise will also adapt with the situation of the moving target, so adaptive tracking of moving target can be achieved.Take the real-time factor into account, we fuse the colour and the vein of the target when we use feature point to establish the model of the vehicles. So we can overcome the drawback when target grey is used as the sole feature to track the target, which is insensitive to vehicles with different colours. By optimizing feature points, we can not only reduce the number of the feature points and improve the match rate, but also increase the total information of the feature points and promote the stability of the tracking process. We provide an image tracking algorithm based on feature point to solve the cover and deformation problem during the tracking process. This algorithm can track the target successfully even when the target is partially covered or has a revolvement deformation. By using matching rate as well as the colour difference between target and background, we can get the situation determination of the moving target and the update strategy of the background template. So we can solve the cover and deformation problem in target tracking process.Finally, we use real video images to carry out some simulation experiments so as to test the robustness of our algorithm. Well performance is achieved and the effectiveness of our algorithm has been proved.