A disparity map refinement to enhance weakly-textured urban environment data

This paper presents an approach to refine noisy and sparse disparity maps from weakly-textured urban environments, enhancing their applicability in perception algorithms applied to autonomous vehicles urban navigation. Typically, the disparity maps are constructed by stereo matching techniques based on some image correlation algorithm. However, in urban environments with low texture variance elements, like asphalt pavements and shadows, the images´ pixels are hard to match, which result in sparse and noisy disparity maps. In this work, the disparity map refinement will be performed by segmenting the reference image of the stereo system with a combination of filters and the Watershed transform to fit the formed clusters in planes with a RANSAC approach. The refined disparity map was processed with the KITTI flow benchmark achieving improvements in the final average error and data density.