A Supervoxel-Based Random Forest Method for Robust and Effective Airborne LiDAR Point Cloud Classification

As an essential part of point cloud processing, autonomous classification is conventionally used in various multifaceted scenes and non-regular distributions. State-of-the-art methods mostly process raw clouds, using a single as the basic unit calculating features by searching local neighbors via k-neighborhood method. Such tend to be computationally inefficient have difficulty obtaining accurate feature descriptions due inappropriate neighborhood selection. In this paper, we propose robust effective approach that integrates supervoxels their locally convex connected patches into random forest classifier, which effectively improves calculation accuracy reduces computational cost. Considering different types descriptions, divide three categories (point-based, eigen-based, grid-based) accordingly design distinct strategies improve reliability. Two International Society Photogrammetry Remote Sensing benchmark tests show proposed method achieves state-of-the-art performance, with average F1-scores 89.16 83.58, respectively. The successful clouds great variation elevation also demonstrates reliability challenging scenes.