Image Matching And Its Application In Automotive Chassis Foreign Object Detection

Currently public safety has become the focus of world attention, because all kinds of car bomb attacks frequently, on-line monitoring of the vehicle chassis is becoming more and more necessary for the entrance of some important occasions. For the majority of people’s lives and social stability, and also to prevent the use of drugs, it is more and more urgent to prevent the vehicle chassis of carrying dangerous goods. Therefore, systems of intelligent vehicle chassis foreign object recognition play an increasingly more and more important role. With in-depth study of computer vision and pattern recognition, the system is more and more inclined to direction of higher speed and more precise. In the process of image detection, images of the same environment may be different for the difference of acquisition time, location, acquisition mode, the viewing angle, and so on. These images are not suitable for directly detecting foreign bodies, so fast and efficient image matching algorithm becomes an integral part of the study.In this dissertation, various feature extraction algorithms on automobile chassis image were in-depth analyzed. Experiments show that SURF algorithm is suitable after a comprehensive comparison of automobile chassis image feature extraction results. The point extracted by SURF algorithm tends to be more uniform distributed and the number is more suitable for the analysis. Distribution of points is important, as more points will take more time, which is not conducive to real-time detection. SURF algorithm extracts feature points according to the extremism point of Hessian matrix, and carries on the longitudinal comparison in different scale space. By comparing each pixel with the surrounding eight points and nine points of two scale layer, added up26pixels, feature points are extracted; and the similarity between feature points is determined according to the main direction and a64-dimensional feature descriptor. The experimental results show that the proposed algorithm to extract the feature points of distribution is uniform, and the number is suitable for analysis. Few points can carry little amount of information, which will make the analysis difficult. After feature extraction,64-dimensional feature descriptors will be computed based on Haar wavelet transform in response to the x direction and the y direction. The matching principle is minimum Euclidean distance for two direction match. Compared with the single direction methods, two-direction matching can significantly improve the accuracy of the match. To further improve the accuracy of the match, the SURF algorithm together with RANSAC algorithm is proposed. Experimental results show that, the phenomenon of mismatch reduces effectively after combining the two algorithms. Finally, the detected image is corrected according to the matching result of standard image and detected image. After distortion correction, the position and size of a foreign object is calibrated. Rapid detection of foreign bodies can be achieved after all the progress.After detecting the position of foreign objects, the depth information needs to determine. Currently, there are a variety of ways to get depth information. In this dissertation, binocular vision is mainly described to obtain depth information. In binocular stereo vision, image matching is very important. SURF algorithm is used for stereo matching, and finally depth information of foreign body is obtained.During the study of depth image, Gaussian curvature and mean curvature is inevitably involved. Taylor function which has shift characteristic is proposed to be used as basic function of moving least squares surface fitting. After fitting we can easily calculate the surface characteristics of each pixel, which lay a foundation for future analysis.