A Robust Structure-Aware Arc-Length Parameterization Framework for 3D Lane Reconstruction

Accurate and structurally coherent 3D lane geometry is fundamental for autonomous vehicle (AV) high-definition (HD) mapping, supporting vehicle localization and planning. Lanes form the structural backbone of road geometry, constraining vehicle pose estimation and guiding navigation. For practical deployment, lane representations must maintain geometric structure and positioning fidelity under degraded or uncertain sensing conditions. Conventional 2D-to-3D lane lifting methods struggle under real-world sensing challenges such as sparse returns from the light detection and ranging (LiDAR) sensor, camera calibration drift, and environmental degradation. Learning-based reconstruction approaches can compensate through learned priors but are prone to sensor overfitting and domain shift degradation. This work introduces a geometry-driven lifting approach that reformulates lanes as arc-length-parameterized curves, producing metrically consistent and structure-reserving 3D polyline representations from monocular 2D detections. This arc-length reparameterization aims to preserve structural coherence under real-world sensing constraints. Sparse LiDAR points are used as depth anchors when available. Depth matches, Random Sample Consensus (RANSAC) plane fitting (on the original point set), camera intrinsics, and the estimated ground surface are fused with structure priors to yield a 3D parametric curve representation of lanes. 2D detected lane anchors are reparametrized as cumulative arc-lengths. A piecewise cubic Hermite interpolating polynomial (PCHIP) maps arc-length to progression along a 3D parametric curve, thus generating 3D lane points and 3D lane structure. The pipeline is modular, lightweight, and non-learning-based, making it suitable for resource-constrained deployments or domains lacking stable priors. Overall, this work establishes arc-length-parameterized geometric lifting as an effective, interpretable, and deployment-ready solution for monocular 3D lane reconstruction, offering strong structural fidelity with minimal sensor requirements and without reliance on learned priors.