Automatic Point Cloud Classification of Under-Represented Pole-Like Objects Based On Hierarchical Directed Graph

Pole-like objects represent important street infrastructures for road inventory and road mapping. Existing supervised point cloud classification methods cannot correctly classify underrepresented pole-like objects in airborne laser scanning (ALS) or photogrammetric datasets due to the limited number of the annotated points. In this article, we proposed an unsupervised method to overcome the challenge of automatically extracting pole-like objects using point clouds. Firstly, the non-ground scattered points are segmented into meaningful segments. Then, DBSCAN clusters are generated form the layered points as the nodes for hierarchical directed graph construction. Finally, a graph-based connectivity analysis in combination with depth-first search is proposed to count the number of directed edges and extract pole-like road furniture candidates. The proposed method has been tested on the Hessigheim 3D dataset at segment- and point-scale. The precision of segment and point scale reached 81.08% and 96.84%, respectively. The experimental results demonstrated that our method can automatically extract pole-like objects robustly and efficiently.