Abstract
Accurate metric depth recovery and camera calibration are essential for 3D perception in traffic monitoring systems, yet most roadside surveillance cameras operate without calibration targets or depth sensors. Recovering consistent metric geometry from monocular imagery in such settings is inherently ill-conditioned due to scale ambiguity and limited geometric observability. This paper introduces a practical hybrid constraint optimization framework for training-free camera calibration and metric depth alignment from static roadside monocular images. The method jointly refines camera intrinsics and affine depth scaling parameters, leveraging road planarity and geometric consistency constraints by combining learned monocular depth, semantic road segmentation, vanishing-point geometry, and vehicle keypoint priors within a robust nonlinear least-squares formulation with a problem-aware multi-start strategy. The approach requires no retraining, calibration targets, or LiDAR supervision. Experiments on the DAIR-V2X-Seq infrastructure-camera benchmark demonstrate consistent quantitative improvements. In a dedicated cross-sequence calibration transfer experiment, the proposed framework improves metric accuracy in over 70% of sequences, with an average RMSE reduction of 3.1m and a δ1 increase of 0.30. Across 28 sequences, per-frame intrinsic estimation exhibits a median focal-length coefficient of variation of 11.3%, highlighting the instability that sequence-level optimization is designed to address. Overall, the study shows that explicitly coupling learned depth with geometric and semantic constraints can improve metric 3D perception in existing uncalibrated roadside camera networks.

This publication has 29 references indexed in Scilit: