Study On The Traffic Flow Detecting System Based On Video

Today, with the rapid development of the economy, the increasing number of cars has put the road transport under tremendous pressure. Therefore, to build a more perfect intelligent traffic system is imminent. The extracting, processing and analyzing of traffic information are very important to the intelligent traffic system. The research of traffic flow detecting Based on video is an important research topic in intelligent traffic system. Based on the analysis of existing algorithms, this thesis makes a research on the vehicle detection, tracking and statistics in real traffic scene. The main works of this thesis as follows:To detect the moving vehicles, a method for the detection of moving object based on probabilistic classification is used. By means of the luminance vector of each pixel for sequences of consecutive frames, background and foreground probabilistic models defined for each pixel are constructed. Using these models, we can calculate the prior probability posterior probability in the foreground or in the background. Finally, according to these probabilities, each pixel is classified in the background or foreground, and the vehicles detection is achieved.For moving vehicle tracking, the thesis made a further study on a object tracking method based on particle filtering. Particles are grouped into clusters using the K-means algorithm. It is necessary to use the background information and convex region formed by the clusters to split more than one clusters associated to only one vehicle and merge one cluster associated to more than one vehicle. These procedures are used to ensure that each cluster is associated with a single moving object in the scene. Vehicle tracking is performed by the similarity between color histograms computed for each of the particles associated with vehicles in the previous frame and in the current frame.The vehicle counting algorithm based on virtual line inevitably exists the possibility of missing and erroneous. Concerning this issue, this paper extract and combine two types of image information, the virtual lines' relative positions with the objects and its changes of waveform of pixel value, then a new vehicle segmentation and counting method was proposed. First, determine the relative positions between the objects and the virtual lines, and combine with the variance of virtual lines' pixel value. With these information, we can improve the accuracy of the traffic flow by means of dividing vehicles.A testing system has been developed using Visual Stdio 2010 and OpenCV for testing the performance of the method studied in this thesis. The system has run in some kinds of weather, and it's result has been analyzed. The results show that the method has excellent performance both in real-time and accuracy. The accuracy was above 95% for each lane of traffic.