Automating construction of road digital twin geometry using context and location aware segmentation

Abstract

Geometric Digital Twins (GDT) represent a critical advancement in road management, yet their practical implementation encounters a substantial obstacle due to development costs outweighing the expected benefits. This paper addresses this challenge and introduces an automated solution for creating 3D geometric foundation models for road digital twins. The proposed approach utilises point clouds to generate meshed, coloured, and semantically labelled models of road objects. The proposed solution incorporates context- and location-aware segmentation, followed by a 3D representation step via meshing. Experiments showed that the solution achieves a 91.7% mean intersection over union segmentation on road furniture in the Digital Roads dataset and surpasses the current leader on the KITTI360 dataset by +16.93%. As a result, the fully automatic method enables scalable and affordable geometry digital twinning for roads.