Estimation of Planar Surfaces in Noisy Range Images for the RoboCup Rescue Competition

Abstract

In this contribution we present a comprehensive method to extract connected planar polygons from a range image acquired by a laser range camera. The result is a boundary representation of the objects in the scene. The approximation of the detected planes by three–dimensional polygons can be used as a source for feature–based registration of sequential range images. We focus on the processing of range images that are segmented into planar regions. The polygon of each plane is extracted by incremental line fitting on the 2D contour of the segmented region projected onto the xy–plane followed by the propagation to the corresponding 3D–plane. We present a novel idea for joining these three–dimensional polygons. Due to sparsely distributed depth values of inclined planes and noise in areas of object edges, some planes cannot be segmented completely. Therefore object edges that actually represent one edge drift apart in 3D. Such edges are detected and joined. The direction in which each edge is moved, is determined by a confidence measure, depending on the slope of the plane. We describe our experiments on range data of scenes containing planar as well as curved surfaces and give quantitative results. The estimates are compared to measurements that were manually taken. For cubic objects we compare the angles of the estimated polygons to the expected orthogonality. For approximately 70 planar surface patches we get an average error of 5 degrees. 83 percent of successfully segmented lines are joint correctly by the algorithm.