Reconstructing Surfaces from Unstructured 3D Points

Most active and passive range nding techniques yield unstructured and generally noisy 3D points. In order to build useful world representations , one must be able to remove spurious data points and group the remaining into meaningful surfaces. In this paper, we propose an approach based on tting local surfaces. Diierential properties of these surfaces are rst used iteratively to smooth the points, and then to group them into more global surfaces while eliminating errors. We present results on complex indoor and outdoor scenes using stereo data as our source of 3D information.