Abstract: Streetlights are critical components of urban infrastructure,timely and accurate acquisition of streetlight information is essential for the development of digital cities. Constrained by the complex object structures and occlusions in urban environments,traditional streetlight extraction methods still suffer from low accuracy,inefficiency,and poor robustness. Additionally,these methods lack general applicability across different urban scenarios. To address these issues,this paper proposes an automatic method for extracting urban streetlights based on vehicle-borne laser point clouds. Firstly,a cylindrical spatial neighborhood is established using ISS keypoints,and potential pole-like objects are identified through density threshold discrimination and back-projection. Then,non-target pole-like objects,such as street trees,are rapidly eliminated using PCA principal vectors,normal vector directions,and angular constraints,resulting in a candidate set of streetlight points. Finally,leveraging the spatial geometric features of streetlight point clouds,a decision tree model is instantiated via a random forest algorithm to match and classify the candidate streetlights,achieving precise extraction of streetlight point clouds. Experimental results indicate that the proposed method attains high extraction accuracy and robustness when dealing with regularly distributed or partially occluded streetlight point clouds,demonstrating significant practical application value.

Key words: vehicle-borne laser scanning, streetlight point cloud, pole-like features, principal vector, random forest