| Stereo matching is a technique that recovering depth information of real scenes from the planar images.The method is to find out matching point pairs from two or more images of the same scene,and then calculate the depth value of the spatial physical point corresponding to the point pairs according to the principle of triangulation.The stereo matching system simulates the principle of human visual perception.It only requires two or more cameras to take pictures of the same scene,and the obtained image can be put into use after stereo correction,which has the advantages of simple implementation and low computation cost.It has a wide range of applications in modern society,such as driverless cars,scene restoration,virtual reality,and so on.However,in the actual matching scene,it is still a huge challenge to achieve high-precision depth estimation through stereo matching due to the influence of external environmental noise,mutual occlusion of objects,the existence of depth discontinuities and low texture areas.Based on the in-depth research on the existing stereo matching algorithms,this dissertation mainly works from the following two aspects:First of all,considering that the lighting conditions of two images to be matched are not always the same,a new algorithm which applicable for various lighting conditions is proposed in this dissertation,called as the pixels' matching cost calculation method based on rank features.It is proposed based on the assumption that the relative order of the local pixels' gray values will not change.The algorithm can obtain the rank order of each pixel in the feature window to reflect the relationship of these pixels' gray values at a low computational cost,and construct the rank matching features that are insensitive to the illumination change.It solves the problem that the relationship of all pixels' gray values cannot be compared in the previous methods,and at the same time,the gradient information of the image can be easily used.These make the matching features contain rich structural information of the image block,and finally obtain the matching cost that can reflect the dissimilarity of two pixels in the image more accurately.The algorithm we propose can achieve high-accuracy matching result whether the matching images have the same lighting conditions or the matching images have a large difference in illumination.Secondly,considering that all existing pixels' matching cost calculation methods have a limited scope of application,a new algorithm that can be adapted to a larger application range is proposed in this dissertation,called as the pixels' matching cost calculation method based on multiple matching cost fusion.The algorithm uses three existing pixels' matching cost calculation methods with different relative merits and applicable scopes as benchmark algorithms.And then it fuses them.For the two different pixels,different benchmark algorithms will first calculate their matching cost and their matching confidence.Then the new algorithm inputs the matching information obtained by these benchmark algorithms into a random forest classifier,and allows the classifier to make analysis decisions and predict true matching cost of that two pixels.In this way,the advantages of three benchmark algorithms are effectively integrated.In a variety of matching scenarios,the new algorithm can achieve higher matching accuracy rate than the benchmark algorithm,which means that new algorithm has a wider application range. |
