Segmentation of Very Sparse and Noisy Point Clouds

Abstract

This paper summarizes an approach to segment 3D point clouds into drivable ground, obstacles, and overhangs. It was developed for outdoor Time-of-Flight cameras which only provide very sparse measurements. The proposed methodology takes advantage of the matrix-like data structure of the CMOS sensor for segmentation in order to increase efficiency. Furthermore, it was tailored to handle typical offhighway characteristics with different slopes and missing measurements and can be adapted to various mounting positions and vehicle properties. First, the algorithm processes the data column-wise using geometric relations. Afterward, the neighborhood of a measurement is considered to improve the initial classification. Finally, overhangs are separated.