A Study On Matching Of Objects With The Same Feature For Stereo Vision

This paper mainly focused on the matching of object that with the same feature under stereo vision. Matching of corresponding points is a evergreen topic in field of image processing and computer vision, including image matching and stitching, target recognition and tracking, target classification and retrieval, image understanding and fusion, threedimensional reconstruction. It hasn't been solved after decades of research. The matching of the objects with the same feature is more universal because they don't have features that vary from application.This problem specified in the bomb-spots matching in volley mode with the system of remote image monitoring and automatic evaluation. The so-called volley mode means that one training team launches several bomb simultaneously or within a limited time. In order to achieve the coordinate positioning of the landing spots, we should match those bomb-spots in both of the images.Although we talked about the problem based on binocular stereo vision, it can applied to multi-vision. The bomb-spots that obtained by the cameras in volley mode distribute in sequence images and have the same feature. Analyzing all kinds of constraint condition of corresponding points matching in binocular stereo vision, We get the solution that based on epipolar constraint and frame synchronization with the assist of GPS.The precision of the epipolar constraints is determined by the estimation accuracy of the fundamental matrix. We explored two methods to find accurate fundamental matrix that suit for the system: taking advantage of the system calibration parameters to get fundamental matrix; utilizing matched points that we have already known to obtain the fundamental matrix. The influence of simplification to the fundamental matrix in the progress of system calibration parameters was analyzed in this paper.The key point in fundamental matrix estimation with matched points is how to acquire a certain number pair of matched points. The image quality of the target area from infrared camera has massive blob and clear feature, we applied a blob-like feature detector and descriptor(ASIFT) to match the background images from infrared camera. Making use of the matched points to get the estimation of fundamental matrix. Adding infrared filter on the visible light camera makes the whole image darker and operating in long distance makes image blur and having greater noise. We applied a strategy of point set registration based on Gaussian mixture model to obtain matching points, Canny algorithm is used to extract image edge which will be used as input in the process of point set registration. In some scene, edge information is too limited to get enough point set using for matching, in the case we employed man-made control points to gain the corresponding points to estimate fundamental matrix. Redundancy of control points was analyzed, meanwhile, cross validation method is applied to get a proper estimation result with limited control points.This paper introduced the image processing algorithm of front-end measuring station and bomb-spots points matching course in host computer. The image processing algorithm including the judgment of the start of trigger, detection of the bomb-spot points, identification and removal of the false target and continuous frame judgment based on image correlation detection of temporal sequence. The points matching course in host computer is based on epipolar constraint and frame synchronization of time from GPS. First of all,utilizing the time from GPS to fulfil frame synchronization. Secondly, Getting the fundamental matrix with different estimation methods according to the number of corresponding points. Then using the time information to deal with the condition that there is only one bomb-spot in both images at the same time. Finally, taking advantage of epiolar and sequence constraint to cope the situation that both images have multi-points simultaneous.