Structure-Aware Calibration Refinement for Monocular Roadside Cameras Using a Horizontal Ground-Coordinate Objective and Road-Patch Phase Correlation.

Monocular roadside cameras require accurate image-to-ground mapping and correction of view changes caused by mounting-structure deformation. We combine a horizontal ground-coordinate calibration objective with road-patch phase correlation and affine correction on selected approximately planar road sections with an approximately constant grade. The initial calibration uses 190 reviewed image-global navigation satellite system (GNSS) correspondences from 14 equally weighted cameras at seven sites. With all available control points used for fitting, the camera-balanced mean horizontal position error (MHPE) at those points decreases from 0.259 to 0.196 m (24.4%), and the all-point root mean square error (RMSE) decreases for every camera. This is in-sample fitting performance at the same control points used to estimate the mapping, not independent predictive accuracy. Complementary landmark-level leave-one-out (LOO) evaluation measures prediction at withheld points: the interior LOO decreases from 0.331 to 0.318 m and the exterior LOO from 0.445 to 0.384 m. Overall LOO decreases from 0.398 to 0.348 m, with improvement in seven out of 14 cameras, indicating geometry-dependent predictive benefit. On the original 40 correction frames, Phase-Affine has a lower observed MHPE under the present annotations: 0.392 m, compared with 0.462 m for native-resolution patch DISK + LightGlue with direct robust affine fitting; both yield 40/40 finite evaluations. Their measured mean core times are 9.858 and 106.094 ms, respectively. A 92-frame hourly experiment examines six approved patch centers, five sizes, normalized cross-correlation (NCC) thresholds, patch subsets and transformation models, retaining large finite errors and separately reporting missing transforms. Full-image learned features are more accurate than unfiltered 150 px Phase-Affine on these hourly sequences. The results demonstrate improved in-sample calibration fit and a lower measured core runtime for phase-based correction in the evaluated patch comparison. The small observed MHPE difference is interpreted cautiously because single-annotator repeatability was not formally measured. Held-out and sensitivity analyses delimit these benefits rather than establish uniform superiority across cameras, patch configurations or operating conditions.