Research On Cross-border Vehicle Rapid Identification Method

With the development of economy, the numbers of vehicles in each country are growing quickly. Many countries pay more and more attention to the intelligent transportation system because of the flourishing traffic. Intelligent transportation system can relieve traffic pressure, and provide effective data sources for the intelligent analysis and scientific management to traffic. Illegal cross-border behavior is a common traffic activity in violation of regulations. Detection and identification aiming at the cross-border vehicles are analyzed and studied deeply.At the first, several common image pre-processing methods are introduced. The previous detection methods need to traverse a large number of pixels within the region of interest, or need to track the moving vehicles in the video, which possess some disadvantages such as complex processes, large amount of computation, and even more poor stability of the algorithm. So we do the following research:the extraction of key points is developed based on regional growth, and only small-scale neighborhood of the key points need to be detected, thus the detection velocity has been enhanced, and then the image binarization threshold is determined adaptively according to pixel feature of key point in HSI color model. In addition, the proposed method in the paper is suitable for vehicles of all kinds of color, and also can overcome the influence of changing light.And then, the three specific steps of license plate identification are proposed:license plate location, character segmentation and character identification. In addition, the operating principle of BP neural network is introduced. Many researchers brought momentum factor into BP neural network to overcome the disadvantages of BP neural network, such as slow velocity of convergence and easy to fall into locally optimum solution. But there are still some problems needed to study about the uncertainty on specific value of momentum factor. Experiments and analysis are designed to illustrate the function of momentum factor in detail, as well as to present the effect to training process from specifying the value of momentum factor. Results display that greater momentum factor and training step length should be selected appropriately.Finally, the whole detection and identification system is built by using MFC in Visual C++6.0. The algorithms mentioned above are realized, the feasibility of these methods is proved by experiments with actual video.