Abstract: Tree filter matching method will cause wrong matching results since it considers only one single component with color to obtain the weight of support region. This paper presents a local tree filter stereo matching method based on pixel classification and color segmentation. The pixels of an image are classified according to their stability during their initial matching cost computation phase. A tree structure based on color information of reference image is constructed, meanwhile, a segmented image with color segmentation constraint is generated. The weight value of each edge is improved by utilizing the information of segmented color image and classified pixels. Finally, the tree filter is executed and the dense disparity is achieved. The experimental results on Middlebury datasets show that our proposed method has higher accuracy than other original tree filter methods in each special region.