Research On Stereo Matching Technology Based On Belief Propagation

Stereo matching technology generate dense disparity map by relating the image with the scene,and it's very significant for some application fields like 3D TV,machine navigation and virtual reality,etc.Stereo matching technology which based on the belief propagation algorithm can get dense disparity map precisely by spreading belief confidence of a node' marked disparity in the neighborhood,and it is a global algorithm which solve an optimization problem with high matching rate.However,the difficulties of stereo matching can bring about a sharp decline in the accuracy of matching,and the complexity of the algorithm make it difficult to meet the quick matching requirements.This thesis mainly research on how to realize high precision and quick speed matching by the belief propagation algorithm.The main work includes:(1)There are some error matches at depth discontinuity regions in the disparity map,and the problem is caused by the jump of the depth between different objects in an image.So this thesis puts forward a construction method which has adopted smooth prior model in energy function based on belief propagation.First,a prior model is combined with the SAD transform which based on rank similarity measurement method as a new measurement method,then the new measurement method is combined with gradient measurement method as a coalition similar measurement method to calculate the data term of the energy function.Then the calculated reliable disparity is used as a new constraint condition for energy function.Next,an experiment is done to verify the efficient of the improved belief propagation algorithm.Experimental results show that after introducing the smooth prior information,the disparity map has less mismatch rate in depth discontinuities regions.(2)In the process of updating the message of confidence iteratively,the message of some nodes at depth discontinuity regions is wrong and pointless.So this thesis puts forward a method detecting the node at depth discontinuity regions by combining the initial matching cost value curve.For some nodes at the depth discontinuities area,the error spreading of it' message is stopped,as a result,the missing match rate at the depth discontinuities area is improved and the running time of the algorithm is reduced.The belief need to be updated only for some yet convergent nodes,so the calculation of the converged node is prevented,and the efficiency of the algorithm is improved.The optimization is combined with three accelerated technology as the improved belief propagation algorithm,then,an experiment is done to verify the efficient of the improved belief propagation algorithm.Experimental results show that,in contrast to standard belief propagation algorithm,the running time is reduced and the mismatching rate at the depth discontinuities area is decreased by the improved belief propagation algorithm.(3)The improved belief propagation algorithm is applied to an euclidean reconstruction system and the model of the scene with a true scale is extracted,the feasibility of the optimization in this thesis is verified.