Research And Implementation Of Global Stereo Image Matching Algorithm

By comparing the different images of the same field in different locations,stereo matching can obtain the matching points of the two images.The traditional stereo matching is mostly sparse matching based on points feature,but due to its characteristics of high local texture repeat degree,more discontinuous disparity,foreground obscured background etc.,so these defects make local stereo matching algorithm having slow matching speed,high matching error rate,more sensitive to the occlusion area and non-texture area,higher degree of difficulty of image dense matching etc..Compared with the local stereo matching,global stereo matching can obtain high matching accuracy.Therefore this paper selects the global stereo matching based on belief propagation to get the global matching of the two images,the main content of this paper is to do the following two improvements in the basic algorithm based on belief propagation.The Belief Propagation is recognized as a global matching algorithm,which can get a more accurate disparity map,but it needs large amount of calculation and its time complexity is higher.So part of this paper is to reduce the time complexity of the algorithm,and to improve the operating efficiency of the algorithm.The paper cited three optimizations for the Belief Propagation.Firstly,use distance transform algorithm to optimize message calculating,which can reduce the complexity of the update of message;Secondly,use the technology of Bilateral Figure to update the iterative on message,which can reduce half of the amount of iteration calculation of the message,and which cannot effect the calculation accuracy;Thirdly,multi scale belief propagation can speed up the convergence rate;In this paper,combine these three program can improve the operational speed of the algorithm.Meanwhile,in order to ensure the accuracy of matching,this paper achieved the occlusion detection though left and right consistency checking and then exclude the occlusion.Experimental results show that the improved belief propagation algorithm has obvious advantages in terms of speed,and also get a good disparity map results.The traditional belief propagation algorithm is based on the pixels of point,and it not only has a higher time complexity,but also has a higher risk of wrong matching at noise points,so this paper combines image segmentation algorithm and belief propagation algorithm.Firstly,we can obtain the segmentation results and edge information of the two images by Mean Shift segmentation,and sign the edge points.Secondly,we can use the non-local cost aggregation algorithm to obtain the initial disparity value,and then get reliable pixels and unreliable pixels by left and right consistency checking and similarity filtering,then in according to the quantity of the reliable pixels and its proportion of the currently discussed segmentation area,the area can be divided into reliable area and unreliable area,and then the reliable area disparity plane can be fitted though Least Trimmed Squares,approximate unreliable area depending on the reliable area disparity plane pattern,though combining the adjacent area which has same or similar disparity plane pattern we can get the final disparity plane pattern set.In the process of global optimizing of the disparity plane pattern,by improved the formula for the message propagation and belief propagation of the single pixel,using the segmentation region instead of pixels and the disparity plane template set replacing the disparity search area,to make the message transmission between pixels converted to the propagation of message between the segmented regions,and achieve the global optimization of the disparity plane template of each segmented region,then calculate the final disparity of the area.The experimental results show that this method reduces the time complexity of the algorithm,at the same time,the wrong matching for noise pixels is reduced through image segmentation and the performance of dealing with occlusions and texture area is improved,finally we can obtain the dense disparity map.