Title: Reconstructing Work Zone Geometry via Pose-Conditioned UWB Range Denoising in V2I Systems

Abstract: Safe and seamless navigation for connected and autonomous vehicles (CAVs) through construction zones relies heavily on accurate work zone mapping. Cone-mounted ultra-wideband (UWB) roadside units (RSUs) present an economical solution for inferring work zone layouts, leveraging direct vehicle-to-infrastructure (V2I) range constraints established by roadside anchors and vehicle tags. Despite their potential, practical field deployments of UWB ranging often suffer from degraded estimation quality due to non-line-of-sight (NLOS) errors, bursty outliers, uncertainties in vehicle pose, and issues related to arbitrary anchor ordering. To mitigate these obstacles, this research introduces a permutation-equivariant predictive denoiser conditioned on vehicle pose for multi-anchor UWB ranging. The proposed architecture utilizes shared anchor-wise temporal prediction to model range dynamics, symmetric set aggregation to manage unordered or missing anchors, and pose-conditioned residual decoding to integrate vehicle motion as a geometric prior. Training follows a two-stage approach: initial learning from observed ranges, followed by fine-tuning the denoiser using NLOS-weighted supervision. The method was validated using large-scale controlled simulation benchmarks for ablation studies and rare real-world V2I UWB field data gathered from a CAV. Findings indicate that the approach significantly enhances range accuracy, cone localization, and work zone geometry reconstruction, particularly in NLOS-dominant scenarios. Furthermore, the system demonstrates robustness against anchor re-indexing and moderate anchor dropout, achieving a 66.9% reduction in measurement-weighted field Mean Squared Error (MSE) compared to raw input data.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC