Supervoxel-based extraction and classification of pole-like objects from MLS point cloud data

The three-dimensional (3D) instantiation expression of objects in road environments has significant applications in urban management and high-precision mapping. Pole-like objects, as the important part of road objects, their automatic and individual extraction and classification can reduce the cost of mapping and improve work efficiency. Therefore, this paper proposes a supervoxel-based method to automatically extract and classify individual pole-like objects from mobile laser scanning (MLS) point cloud data. First, supervoxels are generated through over-segmentation, and the vertical pole part of the pole-like objects are extracted via supervoxels region growing. Second, an uphill clustering method is used to individually segment the potential attachments on the vertical poles (the part of the pole-like objects except the vertical poles). Third, according to the spatial correspondence between the vertical poles and their attachments, the attachments are selected from the potential and matched with their corresponding vertical poles. Finally, the extracted individual pole-like objects are classified into four categories (street lights, cantilever traffic poles, street trees and others) according to their geometric characteristics. The proposed method was evaluated using two different MLS point cloud datasets. The experimental results demonstrate that the proposed method can efficiently extracted the pole-like objects from the two datasets, with the extraction rate of 92.4% and 98.6% respectively. Moreover, the proposed method can effectively classify the extracted pole-like objects.

Automated summary

This paper proposes a supervoxel-based method to automatically extract and classify individual pole-like objects from mobile laser scanning (MLS) point cloud data. The method efficiently extracts the pole-like objects and can effectively classify them into four categories based on their geometric characteristics. The experimental results demonstrate a high extraction rate of 92.4% and 98.6% for the two datasets, indicating the effectiveness of the proposed method.