Dr Shirin Malihi

Title:  A Quality-aware Road Digital Twin Using Multi-modal Data

Theme: Digital Twin

Abstract:

This project develops an uncertainty-aware road digital twin that fuses multimodal sensing—image-derived and laser-scanned point clouds—to represent road geometry and assets with measurable confidence. Beyond building a detailed 3D road model, the work quantifies uncertainty in the twin’s underlying point cloud and studies how that uncertainty affects downstream road-asset detection. The central aim is to use confidence estimates to improve detection performance and reliability, particularly for safety-critical assets that are rare or appear in complex, co-occurring configurations, including challenging aerial mapping scenarios.

Scientific concept and goals

The project focuses on developing uncertainty-aware road digital twins—digital representations of road infrastructure that not only capture what assets exist and where they are, but also how confident the system is about them. By treating the twin as a probabilistic model, the research integrates multimodal

sensing data (vision-derived and laser-derived point clouds) to build and maintain accurate, dynamic road twins. The central idea is to explicitly model and quantify confidence across both geometry and asset layers, using these uncertainty estimates to enhance the detection and interpretation of safety-critical road assets, particularly in challenging scenarios such as rare objects, co-occurring assets, and

complex or degraded sensing conditions. The overarching goal is to make road digital twins more reliable for safety-relevant decisions, enabling them to attach trust scores to detected assets, flag uncertain regions for verification, and support downstream applications—such as maintenance planning, inspection, and risk monitoring—with transparent and quantifiable reliability rather than blind confidence.

Summary of work carried out

The project began with a literature review to establish methods for building road digital twins from multimodal data, focusing on data fusion approaches, reconstruction quality, and sources of uncertainty.

Image-based 3D reconstruction methods, including photogrammetry, SfM-MVS, and visual SLAM (V-SLAM), were then used to generate point clouds from imagery, enabling scalable passive-sensing 3D capture.

In parallel, laser-scanning and SLAM-based workflows were investigated to generate laser-derived point clouds, providing an active-sensing counterpart for comparison in terms of accuracy, robustness, and performance in complex road environments.

A major focus was uncertainty analysis of both image-derived and laser-based point clouds. The research evaluated how reconstruction uncertainty affects the confidence and reliability of downstream road asset detection, particularly for challenging scenarios involving rare assets and closely co-occurring objects.

Overall, the work aims to improve road asset detection by incorporating confidence estimates, helping identify reliable detections, flag uncertain results for review, and enhance decision-making for safety-critical applications.

Major outcomes and significance

This project establishes a robust framework for road digital twins (RDTs) by integrating image-based and laser-based data to create reliable, high-quality infrastructure models.

Key outcomes

• Multimodal RDT pipelines: Developed and evaluated both image-based (photogrammetry, V-SLAM) and laser-based (LiDAR, SLAM) workflows for 3D road asset reconstruction.
• Uncertainty quantification: Introduced a "trust layer" that measures how factors such as noise, occlusions, and point density affect asset detection reliability.
• Confidence-aware detection: Improved detection of rare and safety-critical road assets by incorporating confidence estimates into the perception pipeline.

Significance

• Supports better decision-making by identifying uncertain detections for review and highlighting areas requiring additional data collection.
• Improves robustness by reducing errors in complex or cluttered road environments.
• Enables evidence-based sensing strategies by providing a consistent framework to compare image-based and laser-based approaches.

Overall, the research advances road digital twins from geometric representations to dependable, reliability-aware tools for infrastructure management and safety-critical decision-making.

Future direction
Future work will focus on turning confidence into a direct control signal for detection and validation inside the road digital twin. The next step is to improve detection methods that explicitly use per-point/per-region confidence. A dedicated emphasis will be placed on rare and co-occurring road assets by combining confidence-driven sampling, class-imbalance mitigation, and scenario-specific evaluation protocols. The twin will also be extended to monitor confidence over time, enabling automated alerts when asset changes are detected