Research On Stereo Matching Algorithms For Image Pairs Of ’Ill-posed’ Scene

Stereo matching as the basis for depth information estimating and completely3Ddescription of natural scene is the core component of stereo vision. Currently，researchers had proposed a number of stereo matching algorithms， includingadaptive-weights local matching algorithm and belief propagation global matchingalgorithm which are thought to be the most representative algorithms among all of therelated literature, as a result of their superior performance on the standard stereoimage pairs. Unfortunately, their performance badly decreased when they wereapplied in actual scene, showing as mismatch in low-texture region, invalidation inradiometric varying area,‘stair-case’ in non-parallel plane and ‘foreground-dilation’when occlusion occurs. The reason for all of these effects which seriously affects theperformance of the algorithm is that these areas are ‘ill-posed’ areas which didn’tsatisfy with the(implied)constraints in traditional algorithms.In this paper, the research is focused on the stereo matching problem for‘ill-posed’ scene. Several excellent algorithms were proposed after the fundamentalreasons for the ‘ill-posed’ scene problems had been dig out.The main work and innovations of this paper are listed as following:(1)Point out the limitations of the traditional matching algorithm in ‘ill-posed’regions after the analysis of the basic theories for stereo vision system and theperformance of typical stereo matching algorithms.(2)Proposed an improved adaptive-window epipolar distance transformationalgorithm, which solved the window sensitive issue of the original transformationwas solved and had significantly improved robustness.(3)Proposed an improved adaptive error-coefficient eipolar distancetransformation algorithm in log-chromaticity color space, which made segmentationratio matching primitive independent of the radiometric condition by logarithmictransformation and adaptively set of the error coefficient. Compare with the state ofart ANNC algorithm verifies the effectiveness of the proposed algorithm;(4)Proposed a novel local stereo matching algorithm based on best slant-plane parameter estimation, in which, the traditional fronto-parallel support region wasinstead with the best slant-plane support region, in the consequence of "star-case"effect had been significantly eliminated;(5)Proposed a new particle belief propagation algorithm with slant-planeparameter smooth constraints, in which the traditional disparity smooth constraintswere instead by slant-plane parameter, resulted in the "stair-case" effective wasfundamentally erased.(6)Proposed a novel occlusion detection based belief propagation algorithm, inwhich the smoothness constraints between occlusion pixels and its non-occlusionneighbors were canceled according to the detected occlusion map, in theconsequence of the ‘foreground dilation’ effect was significantly eliminated.