Deep 3D Segmentation and Classification of Point Clouds for Identifying AusRAP Attributes

摘要

Identifying Australian Road Assessment Programme (AusRAP) attributes, such as speed signs, trees and electric poles, is the focus of road safety management. The major challenges are accurately segmenting and classifying AusRAP attributes. Researchers have focused on sematic segmentation and object classification to address the challenges mostly in 2D image setting, and few of them have recently extended techniques from 2D to 3D setting. However, most of them are designed for general objects and small scenes rather than large roadside scenes, and their performance on identifying AusRAP attributes, such as poles and trees, is limited. In this paper, we investigate segmentation and classification in roadside 3D setting, and propose an automatic 3D segmentation and classification framework for identifying AusRAP attributes. The proposed framework is able to directly take large raw 3D point cloud data collected by Light Detection and Ranging technique as input. We evaluate the proposed framework on real-world point cloud data provided by the Queensland Department of Transport and Main Roads.