Low-Cost 3D Spatial Data Construction in GNSS-Denied Tunnels Using Monocular Driving Video and HD Road Maps
TL;DR
This study presents a pipeline for constructing an HD-map-referenced three-dimensional point cloud in an absolute coordinate system from a single low-cost monocular driving video and demonstrates a practical low-cost route for constructing spatial data in GNSS-denied road sections.
Abstract
Tunnels remain a blind spot in three-dimensional road spatial data because Global Navigation Satellite System (GNSS) signals are blocked, repetitive textures and abrupt illumination changes hinder image-based processing, and survey-grade mobile mapping systems are costly. This study presents a pipeline for constructing an HD-map-referenced three-dimensional point cloud in an absolute coordinate system from a single low-cost monocular driving video. Depth Anything 3 is applied to overlapping frame windows, followed by per-window metric scale recovery and anchoring to the horizontal and longitudinal alignment of the national HD-road-map centerline. In a 1.1-km expressway tunnel, the internal scale of the up-to-scale reconstruction varied by a factor of 6.3, preventing stable long-range chaining. The vision-only metric reconstruction estimated a travel distance of 871 m, which was 237 m (21.4%) shorter than the 1,108 m map reference and required a multiplicative distance correction of 1.272. The resulting HD-map-referenced point cloud contained approximately 0.7 million points in UTM Zone 52N coordinates. Its elevation band followed the mapped longitudinal profile, while an aggregated mid-tunnel cross-sectional envelope reproduced the tunnel arch. The pipeline ran on a commodity CPU and demonstrates a practical low-cost route for constructing spatial data in GNSS-denied road sections.