The Algorithm Research Of Vehicle-logo Recognition Based On Image Registration

In the Intelligent Transportation System(ITS), vehicle feature detection includes vehicle-plate detection, vehicle-logo detection, vehicle-type detection and vehicle-face detection. At present, vehicle-plate recognition has been developed quite mature in every country and has been invested in a wide range of applications. As an important research field, vehicle-logo recognition could be used to monitor the road, manage the parking lot, and do electric-charge for highway by assisting vehicle-plate recognition. It has great significance for the process of science and technology and the development of economy.With the continuous development of computer vision and pattern recognition, the ITS on the basis of image understanding Gradually gets actual application. Computer vision simulates human eyes vision by computer, it extracts the useful information in the image and processes and understands the information to implement the actual test, control, and measurement etc. of the system. As a specific application of ITS, image processing technology is widely applied in the field of ITS.This paper mainly aimed at researching and discussing the vehicle-logo recognition technology in the ITS. The algorithm is generally divided into three parts: image information pre-processing, vehicle-logo location, and vehicle-logo recognition. The vehicle-logo locating part is the key stage of the process, includes rough location and accurate location. It uses the knowledge of relative location between vehicle-plate and vehicle-logo to determine the logo’s general region. Then uses Sobel operator to detect the vehicle-logo’s edge, takes advantage of OTSU to convert the vehicle-logo contour to binary image, by calculating the proportion of foreground pixels in the binary image to determine the type of the interference stripes around the vehicle-logo. Then takes morphology smoothing to get the connected domain meeting some conditions. At last, it takes horizontal and vertical projection for the connected domain to get the accurate location by setting the threshold value; vehicle-logo recognition based on feature description and feature matching is approximately divided into three stages of feature points detection, feature description, and feature matching. First it builds templates feature database, then extracts SIFT features for unrecognized vehicle-logo, then takes advantage of improved-FREAK descriptor to descript features. At last by calculating the distance between template images and unrecognized logo to rapidly matching and optimization combining k-nn matching and RANSAC algorithm.In the vehicle-logo location stage, this paper joins the±45°directions convolution templates, by judging the logo’s texture type to do gradient convolution in various directions so that it filters out the interference stripes around by side of extracting the logo’s edge. It filters disturbed stripes around by side of extracting vehicle-logo edge. In the vehicle-logo registration stage, this paper proposes the method of combing SIFT to detect feature points and FREAK to descript feature points. It not only gets more features information, but also obviously shorten the description and matching time. It adds long-pairs of points to original FREAK algorithm, sets threshold Dmin, uses only long distance points in sampling pattern to generate angle information. So that it is suitable for large-rotation-scale environment. In the matching stage, this paper weights the Hamming distance. For every point, calculates mean value of every column for training data descriptors, the more closer the value to 0.5, the bigger the weight. It has been improved the rough state of the original Hamming distance to more accurate.The experimental results show that, this algorithm has better recognition precision and real-time performance. It is more suitable for the applications with large variation of rotation and scale and high demand of matching performance.