Surface model generation from range images of industrial environments

We present an hybrid segmentation technique that combines both the speed of an edge based approach with the robustness of a surface based approach. It consists of three stages. In the first stage a scan line approximation process extracts the edges contained into the given range image. These edges are later on used to define the positions of seed points. Through the second stage a two steps region growing technique is applied. First a 2D growing process enlarges the original seed points generating bigger regions. Next, each region is fitted to a plane and a cylinder. The one that best fit the given points is selected to represent that region and used during the 3D growing stage. The 3D growing stage is carried out taking into account the approximation error from candidate points to be added to the fitted surface. In this way, each surface is grown until no points can be added according to a user defined threshold. Finally, in the third stage, a post-processing algorithm merges neighbour regions that belong to the same surface. Experimental results by using industrial environments are presented.