Efficient multi-plane extraction from massive 3D points for modeling large-scale urban scenes

摘要

In modeling large-scale urban scenes, extracting reliable dominant planes from initial 3D points plays an important role for inferring the complete scene structures. However, traditional local and global methods are frequently prone to missing many real planes and also appear powerless when massive 3D points are present. To solve these problems, the paper presents an efficient multi-plane extraction method based on scene structure priors. The proposed method first explores the potential relations between the planes by detecting 2D line segments in the projection map produced from initial 3D points (i.e., simplify 3D model to 2D model), including: (1) multi-line detection in regions by the guidance of scene structure priors; (2) multi-line detection between regions under the Markov Random Field framework incorporating scene structure priors. Then, according to the resulting plane relations, a rapid multi-plane generation is carried out instead of the time-consuming plane fitting over 3D points. Experimental results confirm that the proposed method can efficiently produce sufficient and reliable dominant planes from a vast number of noisy 3D points (only about 8s on 2000K 3D points) and can be applied for modeling large-scale urban scenes.