Enriching and maintaining road digital twins with condition information

Abstract

Road inspections are mostly carried out manually, limiting their scalability and creating subjective human errors. Research into automating them was studied, though there are still gaps in how to efficiently replace various aspects of manual inspections, such as how to improve detection accuracy and how to integrate them into a defect progression tracking pipeline. Tackling these gaps would enable a digitalised solution for detecting and registering road defects, such as a road digital twin. This paper aims at closing these gaps by providing two unique contributions: Firstly, performing an analysis of the performances of defect detection using different data modalities (RGB images and point cloud data) under different conditions to outline the strengths and weaknesses of each modality. Secondly, to create a pipeline for using past defect information to guide detection through detecting within bounding boxes of known defects and to enable continuous tracking of defect conditions compatible with the IFC format. Mask RCNN was used for detection in the experiments. Results indicate that incorporating information from different modalities can indeed lead to more consistent and accurate detections, and past defect information can enhance detection accuracy in the case of previously known instances of defects.