HDMap-SLRC: Semantic-Landmark Residual Correction for Monocular HD-Map Georeferencing

Low-cost monocular cameras can reconstruct road scenes in three dimensions, but the scale and local coordinate frame of the reconstruction must be linked to an absolute spatial reference. This study proposes HDMap-SLRC, an offline georeferencing framework combining HD-map centerline-based alignment with semantic-landmark residual correction. The geometric stage establishes a map-referenced trajectory, and a Gaussian-process module estimates conservative along-route corrections from sparse semantic observations. In the primary sequence, traffic-light map-referenced error decreased from 24.81 m to 5.87 m, pole error from 19.78 m to 11.21 m, and the 95th percentile (p95) of road lateral distance from 5.302 m to 2.741 m, relative to the 4-DOF baseline. Landmark metrics were interpreted as map consistency, considering the relationship between correction inputs and the evaluation procedure. Across 28 parameter configurations on two reconstruction variants of the primary sequence, the range of road lateral p95 was approximately 3.0% of its mean. In an additional Yeouido acquisition comprising four segments, 992 m, and 160 frames, the mean of the segment-wise median horizontal trajectory errors against an RTK-GNSS/INS reference decreased from 21.16 m to 10.55 m. A scale-perturbation analysis of the saved geometric stage examined how input-scale changes propagate to aligned positions. The results support the feasibility of the staged georeferencing structure under the evaluated conditions. The remaining meter-level errors do not establish suitability for lane-level autonomous driving, and generalization across weather, illumination, and urban layouts remains unvalidated.