Research Of License Plate Localization Algorithm Based On DDMCMC

As the core part of the intelligent traffic management, Auto Vehicle License Plate Recognition System(LPRS), based on computer vision, is an important research topic, integrating such multi-technology of different fields as digit image processing, pattern recognition, artificial intelligent and so on. At present, part of LPRS has been successfully applied to the actual scenes both at home and abroad, such as highway charge management, bridge charge management, vehicle access control of some special sectors, intelligent parking management and the detection of drive against traffic regulations. License plate localization technology is the key point of LPRS. The accuracy of the location has a direct impact on the quality of all the following parts, thus affects the performance of the entire system.In general license plate localization algorithms, only one of such features as color, shape and characters is used to define the license plate, which is not sufficient for accurately localizing the plate stably in different situations. The proposed algorithm absorbs the advantages of different algorithms. And by analyzing the intrinsic features the license plate, aiming at solving the problem of accurately localizing the license plate of traffic intersection in different situations, this paper proposes its technical routine:First, this paper gets the general area of the license plate by using method of image projection. Then the accurate location of the license plate is got by means of DDMCMC (Data Driven Markov Chain Monte Carlo).There are some good points in the proposed algorithm. First, the localization algorithm combines coarse and accurate localizing algorithm, which improves the efficiency of the algorithm, the detection rate and the accuracy of the localization. Second, the DDMCMC algorithm is modified from the MCMC (Markov Chain Monte Carlo) algorithm, which uses the Bayesian theory as a frame. And to get more accurate location of the plate, features of color, shape and characters of the license plate is mixed together. Third, a kind of data driven model, based on geometry feature of the plate, is proposed to modify the MCMC algorithm, which accelerates the Markov chain convergence. More other the data driven model can be used in other image processing fields to detection rectangular objects. Fourth, when coarsely locating the license plate, this paper uses straight-line linear filter to enhance contrast of the edge of license plate area and non-plate area, to reduce the interference of edges caused by the non-plate region. At last, In consideration of real-time problem, in this paper such methods as integral image and pre-getting the color of license plate are used to accelerate the proposed algorithm.For validating the proposed method, we design four sets of experiments. The first set uses a license plate database from traffic intersection of China. The database contains many license plate images with different weather, illumination intensity, partial occlusion and different colors. And we get 97.4% recall rate and 93.85% precision rate with 95% cover rate, which shows that the proposed algorithm is suitable for different situations with high accurate localization. By comparing with the MCMC algorithm, the DDMCMC modified by the data driven model is more efficient. The second set is a contrast experiment using the common license plate dataset Caltech. The experiment result shows the proposed method can ensure a high detection rate with more accurate localization. The last two sets are traffic sign detection and door detection, which use common dataset Stereopolis and dataset from CUNY respectively. The experiment results show the data driven model has its generality.