Segmentation, outlier detection and feature identification from unstructured 3D point clouds

In many fields of science and engineering, e.g. metrology, reverse engineering or computer vision, best fitting of geometric primitives is an important issue. Due to the progress in sensor technology large data sets will be available in real time with low-cost sensors for a wide range of applications, so that the acquisition of spatial data sets is accelerated and becomes less expensive. As a consequence the various areas of applications for segmentation and feature identification methods will be further enlarged and powerful algorithms to cope with large data sets will be needed. In this paper we present a very robust and run-time efficient method for automated feature detection, applicable to any kind of implicit surface or plane curve. At first the mathematical modelling of the task is described and optimization methods with fresh ideas are shown which solve the problem of geometric fitting. Afterwards the algorithmic structure and main steps of the segmentation, outlier detection and best fitting process are explained. Moreover our algorithms are applied to real point clouds generated by a laser-radar scanner. Excellent results are reached even for noisy data sets and partially occluded features. As a conclusion future work and perspectives of the technology are discussed.