Segmenting correlation stereo range images using surface elements

Abstract

This work describes methods for segmenting planar surfaces from noisy 3D data obtained from correlation stereo vision. We make use of local planar surface elements called patchlets. Patchlets have 3D position, orientation and size parameters. As well, they have positional confidence measures based on the stereo sensor model. Patchlet orientations (i.e., surface normals) provide important additional dimensionality that reduces the ambiguity of segmentation-by-clustering. Patchlet size allows the use of continuity or coverage constraints when segmenting bounded surfaces from depth images. We use a region-growing approach to identify the number of surfaces that exist in a stereo image and obtain an initial estimate of the surface parameters. We refine segmentation using a maximum likelihood clustering approach that is optimised with Expectation-Maximisation. Confidence measures on the patchlet parameters allow proper weighting of patchlet contributions to the solution. We provide experimental results of the segmentation on complex outdoor scenes.