Tree Instance Segmentation in Urban 3D Point Clouds Using a Coarse-to-Fine Algorithm Based on Semantic Segmentation

Abstract

3D point clouds acquired with terrestrial or mobile LiDAR sensors are increasingly used to map urban forests. The segmentation of separate tree instances, i.e., subsets of points representing individual trees, is a relevant step in automatically extracting tree inventory data from 3D point clouds. Various algorithms have been proposed for tree instance segmentation, offering different trade-offs between accuracy, runtime, and robustness against data incompleteness and noise. In this work, we propose a coarse-to-fine algorithm for segmenting tree instances in urban 3D point clouds from terrestrial or mobile LiDAR scanning that combines two existing techniques: (1) the computationally efficient marker-controlled Watershed algorithm and (2) a more accurate 3D region growing algorithm. Initially, the marker-controlled Watershed algorithm generates a coarse segmentation, which is further improved by a Voronoi segmentation-based error removal. Subsequently, the coarse segmentation is refined by the 3D region growing algorithm in areas with overlapping tree crowns and sufficient data quality. In both steps, our algorithm uses the results of a prior semantic segmentation to select suitable markers and seed points. We evaluated our coarse-to-fine algorithm in an ablation study using two mobile LiDAR datasets and one terrestrial LiDAR dataset from three German cities. Our results show that our algorithm outperforms the standard marker-controlled Watershed algorithm in terms of panoptic quality by 3.7, 25.5, and 29.6 percentage points, respectively, while being computationally more efficient than an approach purely based on 3D region growing.