| How to get the distance (depth) information from multi stereo images is the mainstudy of stereo vision. Getting the position difference from two or more images capturedfrom different viewpoints is the kemal problem, the position difference called disparitywhich can be computed through triangle measurement theory. Stereo matching is thefundamental and important approach in application of computer vision, which can beapplied in many fields, such as 3-D reconstruction, robot navigation and automation landvehicle navigation, and so on. It is a morbidity problem under the influence ofdeformations, distortions and occulsions. Binocular vision is the main topic in this thesis,the motivation is based on two reasons, first, the theory of binocular more close to that ofhuman's; second, its simple implementation in real application. The relevant theory andapproaches of binocular stereo matching are described; finally, algorithms and experimentsare presented.The main idea of area-based approach in stereo matching is comparing the similarityofpixels in correlation window; the difficult problem is how to select the windows size. Inorder to deal with the problem, we divide all images into two sets, one called character area,and the other called non-character area. Many redundant computations is unnecessary inmatching process, we analyzed the computation process and eliminates the redundantcomputations. In the application of matching large disparity images, for each pixel, thelarge search scope need to fred the correspondence pixel in the other image, but the searchscope can be reduced in some degree. Based on the strategy, the concept of global disparityestimation is proposed and a new fast area-based approach is presented which reduces thecomputation time and search scope.Image transformation is usually used in image processing. The Walsh transformationand Rank transformation are analyzed in the thesis. The information of Walshtransformation is more than that of gray of image, such as zero-crossing, sign oftransformation coefficients. The transformation of Rank is analyzes the relation of currentsignal and its neighbor signal, its transformation result is the sum of pixels which less thanthe current pixel. In this paper, we research the feasibility of using the Walsh and Ranktransformation characters as match primitivity in replace of gray character of image. Theexperiment results indicate that the approach is feasible and got a better result comparedwith using grey character as primitivity.Moment theory has been widely used in image analyzing and processing, the moment sets computed form image usually includes the global characteristics and provides differentgeometry characteristics. The characteristics of image moments can be applied in manyfields. We apply Tebechbief moment characteristic to match replace of using graycharacteristic as primitives. An area moment has the property of unsensitive to transition,rotation and sealed change of area, the property of area moment is applied in the areamatching approach in our thesis, combined with a new constraint, and a novel area-basedmatching approach is proposed.The vision-based approach is a important research field in ALV, whose key problem ishow to get the segmentation area in real time, traditional segmentation approach, such asregion growth, split-merge, which cost too many times to satisfy the real application,multi-threshold segmentation approach usually applied in real application, but how toselect the thresholds is a difficult problem. We define a smooth function to detect thecharacter area, after the images are characterized, the images are transformed in to bi-valueimages, and as a result, its segmentation and matching work become simple and fast. Theapproach needs not to know the prior information and model of the object, so, it can beused in real obstacle detection.
|
PDF Full Text Request