Segmentation for outdoor urban scenes

This paper proposes a segmentation algorithm for 3D outdoor urban scenes. Recently, many 3D range sensors are being used in various fields. In order to grow 3D range sensors, the necessity of various techniques such as objects detection, tracking, classification, 3D SLAM, etc. For the pre-processing step, segmentation algorithms can improve the performance of these techniques. The segmentation algorithm proposed in this paper is composed with combination of two algorithms: planar region segmentation and non-plane region segmentation. In order to segment plane regions, we propose connected components labeling with area opening based noise reduction and boundary detection that take into account surface and gradient direction and sensor model features. For the non-plane region segmentation, graph based algorithm using range features is introduces for sparse 3D point clouds in this paper. The proposed algorithm is verified using publicly available Velodyne dataset. A comparison of conventional algorithm is also represented.