The Research Of Stereo Matching Algorithm Based On Tree Structure

Stereo matching is one of the significant problems of computer vision.The core purpose of stereo matching is to find two corresponding pixels,obtain the disparity information and compute the depth information.Stereo matching is divided into global methods and local methods.The former requires an energy function to iterate while the latter mainly devotes to optimizing support windows and support weights.Recent years,several tree structure methods had been proposed.These methods construct reference image as a spanning tree structure,and achieve cost aggregation by corresponding tree filter.These methods get rid of local support windows,meanwhile,they employ tree structure and tree filter to obtain precise dense disparity map while the complexity of computation is extremely low.However,they may cause the disparity propagation that breaks boundaries while the weight function is not stable as well.Besides,the post processing of tree structure methods show insufficient because of the lack of stable pixels.Aiming at the different weaknesses mentioned before,this paper proposes three modified methods.First,this paper proposes a tree structure method based on region segmentation for stereo matching to solve the problem that the disparity propagation breaks boundaries.We employ a segment constraint to divide reference image as many regions with similar size to avoid the generation of huge planar support regions and eliminate the influence of the disparity propagation that breaks boundaries.Then,to enhance the stability of tree structure's weight function,this paper propose an enhanced tree structure method using modified epipolar distance transform(M-ED Transform)for stereo matching.This method employ both M-ED transform and color information as component to compute the weight function and obtain a tree structure more stable and preciser,so the accuracy of stereo matching promotes naturally.Finally,this paper proposes a post proposing method of tree structure based on region voting of stable pixels.Pixels are divided into stable,unstable and mismatch.Tree filter will be execute after the proportion of stable pixels promote while region voting.The experiment of this paper compares our methods to corresponding classical methods by the data-sets from Middlebury.Experiments prove that our methods perform better in accuracy and stay stabilization.