Evaluation of clustering techniques for segmentation of LIDAR point clouds

Abstract

LIDAR point clouds are helpful for industrial robots, autonomous vehicles, and robotic urban search and rescue because they provide precise distance information. These data sets show indoor and outdoor scenarios with automobiles, people, buildings, and other objects. The information is presented densely and unstructured in 3-dimensional Cartesian coordinates. One of the first tasks in the analysis of the environment shown is the segmentation of the situations. This study evaluates various unsupervised clustering techniques for segmenting LIDAR point cloud data sets. An examination of the classes visually and the computation of clustering metrics serve to examine the behavior of the K-means, DBSCAN, BIRCH, and Mean Shift algorithms in various indoor and outdoor scenarios.