Using Road Design Priors to Improve Large-Scale 3D Road Scene Segmentation

Abstract

Implementing geometric digital twins for roads can increase productivity and help streamline predictive maintenance. However, the related costs outweigh foreseen financial gain due to the manual work needed in scene recognition and modelling. The foremost step in automatically creating such digital twins is segmenting objects in 3D point clouds. Researchers have done extensive work to tackle this challenge, with deep learning solutions often preferable for solving multi-class problems. Yet, only several large-sized road objects get high-quality detection results. Other essential but rather smaller objects represent a further challenge. We utilise common properties of roads and road assets in a lightweight deep learning network training and data pre-processing to facilitate a broader range of highly detected objects. We show how these priors improve detection performance by a significant margin. Our work contributes to the improved automatic detection of most road object types, enabling automatic digitisation for roads and reducing related costs.