Ground Extraction from 3D Lidar Point Clouds with the Classification Learner App

Ground extraction from three-dimensional (3D) range data is a relevant problem for outdoor navigation of unmanned ground vehicles. Even if this problem has received attention with specific heuristics and segmentation approaches, identification of ground and non-ground points can benefit from state-of-the-art classification methods, such as those included in the Matlab Classification Learner App. This paper proposes a comparative study of the machine learning methods included in this tool in terms of training times as well as in their predictive performance. With this purpose, we have combined three suitable features for ground detection, which has been applied to an urban dataset with several labeled 3D point clouds. Most of the analyzed techniques achieve good classification results, but only a few offer low training and prediction times.