Research On Moving Object Detection And Tracking For Intelligent Traffic Monitoring

As the development of intelligent traffic is playing an important role in modern traffic, the study of key techniques of intelligent traffic based on computer vision technology is of great significance. The moving object detection and tracking involve widely different subjects, which was carried out in this paper based on the video processing monitoring of Electronic Toll Collection System in Wuhan city. Taking the actual needs of the development of Wuhan city traffic into consideration, a research in vehicle and pedestrian detection and tracking was conducted, and an algorithm in movement of the vehicle detection and tracking was proposed, which then was modified in specific project to satisfy the requirements of fast development of intelligent traffic. Related experiments was devised to test and verify the feasibility and effectiveness of the algorithm, and the specific research contents are listed as follows:(1) Pair frame difference is used to solve the existing problems of traditional frame difference method。the result of frame difference is kept as the vlaue of the image pixel's value, then the histogram is matched with positive or negative respectively, thus the coarse position of moving targets is calculated; After that, according to the dual relationship of moving objects, which mainly refers that the residual area should be equal, residual comparison is applied to accomplish the goal positioning and finally detect the targets. Through the experiment we can see that this algorithm is of great advantage in detecting the objects with low speed and no shelters.(2) As the results of the pair difference algorithm, the information has large redundancy. The dual edge finite difference method can help reduce the redundancy of the edge information obtained from the frame differential method, which at the same time can guarantee the effectiveness of the testing results and the efforts in shadow removing and in detecting fast moving target.(3) Moving object detection in the analysis of the characteristics of image block is based on the known image information, as the pair finite difference method or dual edge difference method gets only partial information of the target, the graph cut algorithm search object features by setting source and sink, by strongly sampling with simple pair difference method, the reserved information can be guaranteed to be parts of the background or moving object. Then the part-target information is set to be source and the part background is set to be sink, through push-relabel method and setting the pixels own punishment and neighborhood punishment, the source and sink of minimum cut search is complete, the result shows our method can bette locate the target area and achieve goals division.(4) Graph-cut with the overall image cut information consider its ownership in a certain prospect, the whole information can cover the local information in some degree. Cellular automata make the cellular evolution rules to find the moving object feature well based on the cellular neighbors' state, pixel set is established and cellular evolution rules is set.From the pixel set, the moving object's time and space attributes is used to segment the imgae, a good result is relatively excellet with its time and space characteristics. Experiments show that this algorithm which used in target detection is good.(5) A method based on cloud model with histogram of oriented gradient is presented. A region which contains the moving object is found using pair frame difference method. A HOG feature map is established of the region, and then cloud model is used to give and rise the expression of the concept, for the concept of cloud model is qualitative and quantitative combined and the expression is simple, in view of the background and moving object similarity, the algorithm is also strong detection.(6) Mean shift algorithm based on HOG feature is used to track moving object. A fetaute hierarchical model is constructed, which is coarser in whole moving object layer, part based model components features accurate and tracking the moving object in different layers, the moving object use nuclear tracking in two layer is more accurate than the moving object tracking with one layer information, and partially and whole occlusion target location can better be forecasted, and has a good tracking performance after the moving object appearing again.