A Practical Stereo System Based On Regularization And Texture Projection

In this paper we investigate the suitability of stereo vision for robot manipulation tasks, which require high-fidelity real-time 3D information in the presence of motion. We compare spatial regularization methods for stereo and spacetime stereo, the latter relying on integration of information over time as well as space. In both cases we augment the scene with textured projection, to alleviate the well-known problem of noise in low-textured areas. We also propose a new spatial regularization method, local smoothing, that is more efficient than current methods, and produces almost equivalent results. We show that in scenes with moving objects spatial regularization methods are more accurate than spacetime stereo, while remaining computationally simpler. Finally, we propose an extension of regularization-based algorithms to the temporal domain, so to further improve the performance of regularization methods within dynamic scenes.