Multi-camera Cooperative Motion Vehicle Recognition Based On Support Vector Machine

With the rapid development of social economy and the great progress of modern science and technology, the intelligent monitoring system also ushers in the new opportunities of rapid development, now has been widely used in public security, transportation, military, and many other related fields. However, at this stage the intelligent monitoring system still exists many key technical problems need to be solved, such as shelter, real-time problem, etc., especially in the case of traffic congestion or traffic peak, the problem becomes more complicated, making multiple-target detection, recognition and tracking becomes more difficult to achieve. All these problem greatly increase the system requirements for high performance processing algorithms. Due to the single perspective of traditional single camera surveillance systems, it is difficult or impossible to deal with the identification problem in case of occlusion problem. In addition, the recognition rate of single-camera is very low. Response to the above problems, this paper uses the following methods to start analysis and research:Firstly, response to the low update rate and sensitivity to noise problem of the Gaussian mixture background updating model(GMM), we study on modeling the background with a new method combined Support vector machines (SVM) with Gaussian mixture method. This method firstly establishes the initial foreground and background model with Gaussian mixture model. And then, using a sliding template window to scan the current frame image, the previous frame image and the background image obtained by GMM, and calculating the corresponding statistical characteristic value for these three frame image, dividing the foreground image pixels with SVM. In the end, we can obtain the final target detection result by fusing the two-step segmentation result. This method can reduce the effects of noise and improve the accuracy of target detection.Secondly, using a vehicle identification method based on the moving object extraction. This method firstly extracts the moving targets area with a background subtraction. And then, using the SVM classifier based on feature fusion to identify the area extracted before. This method can greatly reduce the detection area, reducing vehicle identification time while reducing the background interference and improving the real-time of the system;Finally, to solve the occlusion problem, we use a moving vehicle identification method based on target matching and multi-camera collaborative. Firstly, each camera independently detects and identify moving vehicles, and then the region matching method and D_S evidence theoretical are used to fusion the recognition result of the same target under different cameras, by this way we can make an optimal decision. This method improves the overall recognition rate of the system.