Research On Automaticvehicle-logo Accurate Localization

Vehicle recognition technique is one of the key techniques in the intelligent transportation system(ITS). Vehicle logo recognition(including logo localization) is a new direction to recognize vehicles and has important application in military logistics vehicle management. The license plate detection can facilitate the vehicle logo localization. These are the two issues addressed in this research work.In the rough logo localization based on license plate detection, under the RGB color model, two new color ratios are proposed and proved to be color invariant by the dichromatic reflection model. They are used to extract the license plate’s color in the license plate detection. Processing 1078 images, the detection rate is 99.91%. The average processing time is nearly 0.2 seconds. Compared with the common used HSV color model based detection method, the detection rate of the latter is 92.12%. Its average processing time is nearly 0.35 seconds. The detection performance based on this paper’s color ratios is better than the performance of the HSV color model. Observing the results, combining the HSV model to extract blue color, the tow ratios of this paper can obtain better result.Good logo recognition needs good logo localization. To improve the accuracy of locating logos and to shorten the processing time, the algorithm combining the analysis of gradient angle histogram and the adaptive thresholding is proposed in this paper. The analysis of the grille gradient angle histogram is used to further suppress the background. The adaptive thresholding is used to cope with the change of illumination and improves the extraction completeness. In the experiment, 5 different logos and 67 images are processed. The logo localization rate is 100%. Compared with the common used template matching based logo localization algorithm, although the logo localization rate of the latter is 100%, but 28 images have large error(41.79% percent of the 67 images). 53 logos(79.1% percent of the 67 images) are better located by the algorithm of this paper than the template matching based algorithm. The performance and robustness of this paper’s method is obviously better than the template matching based method. When the size of roughly localized image is nearly 260*170 pixels, the average processing time of the template matching based algorithm is about 0.04 seconds. The average processing time of this paper’s algorithm is about 0.1 seconds and can satisfy the requirement of the real application.