Quantitative Uncertainty Analysis of Monocular Point Clouds for Twinning Roads

Abstract. Digital twinning of road infrastructure enables the simulation of road asset conditions and performance, supporting more effective planning, management, and maintenance of road networks. RGB cameras offer a low-cost sensing solution for digital twinning applications. Monocular RGB images, in particular, provide a cost-effective source of data for road surface analysis and inspection. From these images, point clouds can be generated to reconstruct road surfaces and surrounding features.
A key factor in the reliability of digital twins is the confidence in the data sources used, especially the point clouds derived from monocular images. This confidence is closely linked to the uncertainty inherent in such reconstructions. In this paper, we analyse the uncertainty of image-derived point clouds generated using two state-of-the-art depth estimation models: Metric 3D V2 and Zoedepth.
Unlike previous studies that focus mainly on geometric accuracy or semantic segmentation, our approach explicitly quantifies the statistical uncertainty and confidence intervals of monocular reconstructions at the object level, thereby introducing a systematic method for reliability assessment in road digital twinning.
We evaluate these point clouds across five distinct object clusters: pavement, tree trunks, tree crowns, lamp posts, and curbs. The analysis leverages a set of geometric features to characterize and differentiate the clusters. Confidence levels for the extracted clusters are quantified using statistical techniques based on normal distribution modelling, allowing us to assess the reliability of classification outcomes.
Our findings indicate that the confidence levels of the identified clusters, extracted from monocular point clouds using a combination of geometric descriptors and supervised machine learning techniques range between 66% and 91%. These results demonstrate that, despite the inherent limitations of monocular depth estimation, meaningful object-level segmentation is achievable with reasonable certainty. Furthermore, the variability in confidence levels highlights the differing levels of geometric distinctiveness and structural complexity among object types.
These results not only demonstrate the practical viability of monocular depth estimation for road asset modelling but also establish a quantitative baseline for uncertainty propagation in data-driven digital twins.