Research On Vehicle Detection And Tracking In Dynamic Traffic Scenes

With the development of electronic technology, in need for regulation of road safety, the density of surveillance cameras installed on roads is increasing, can provide a lot data through video frames for Data Mining. If traffic information can be extracted from the surveillance video, can increase the density of traffic information collection, enhanced the effectiveness traffic information and make it become more real-time, has high value in application.Traditional vehicle detection systems based on video generally detect and track moving vehicles under static scenes, in contrast, surveillance cameras can be rotated and stretch as needed to monitor different areas, which will cause scenes change dynamically. Therefore, this paper research target detection and tracking of dynamic scenes. Therefore, research on methods to detect and track vehicle targets in dynamic scenes.A method based on the accumulated sparse optical flow is proposed to achieve lane positioning and the use of a method based on morphological filtering is presented to extract the lane. The accuracy of this method of locating lane is high, can effectively recognize the background region and foreground regions, and has strong anti-interference capability. On the basis of lane positioning, according to the optical flow field of the background area, a method for determining scene’s state of motion and static is proposed.In dynamic scenes, a method of motion compensation using affine transformation mechanisms is proposed, and Gaussian Mixture Model (GMM) is used to detect and identify vehicles, after apply the morphological methods, information about vehicles’ size and location can be extracted. On this basis, the Camshift algorithm based on Kalman prediction is used in tracking vehicles. In dynamic scenes, the search accuracy of window positioning based on Kalman prediction is decreased, a method of motion compensation is proposed to correct the position of predicted search window, the search efficiency is improved.A prototype system is developed based on OpenCV library and Visual Studio 2008, with VLC library as the decoding tool, and DirectDraw as the graphics engine. Results in experiments show that the prototype system can achieve high accuracy in positioning lane, in the case of dynamic scenes, the system can perform well in detecting and tracking vehicle targets, and basically be real-time.