Decoupled Semantic Registration for Surround‐View SLAM: High‐Precision Mapping and Robust AGV Localization

ABSTRACT

Reliable mapping and localization are crucial for automated guided vehicle (AGV) navigation in industrial environments, where ground markings provide stable, infrastructure‐free landmarks. However, dense region‐based registration in surround‐view semantic simultaneous localization and mapping (SLAM) can be insensitive to small inter‐frame misalignments within overlapping regions, allowing residual along‐marking misalignment to persist and accumulate during mapping. To address this, we propose a decoupled semantic registration SLAM (DSR‐SLAM) framework that separates the geometric requirements of offline mapping and online localization. During mapping, edge‐based registration uses semantic edges to make correspondence errors more sensitive to inter‐frame misalignment, thereby reducing residual along‐marking misalignment and improving pose estimation. During online localization, region‐based registration exploits the fault tolerance of dense semantic regions to maintain robust tracking despite marking wear, fading or partial occlusion. A four‐camera surround‐view module provides a full ground‐level field of view (FoV) for perceptual redundancy. Experiments on two self‐collected industrial datasets and the public AVM‐SLAM dataset demonstrate improved mapping performance over representative SLAM methods, including a map‐distance RMSE of 0.465 m on the public dataset. The proposed framework achieves mapping accuracies of 2.20 and 4.02 cm and mean localization errors of 3.07 and 4.65 cm on the two industrial datasets, respectively.