Dynamic-Aware Point-Line Visual SLAM with Loop-Consistent Semantic Mapping

Abstract

Simultaneous localization and mapping (SLAM) plays an important role in autonomous vehicles and Unmanned Aerial Vehicles (UAVs) when satellite navigation is unreliable or unavailable. However, its performance often degrades in complex scenes with moving objects, weak texture, and map inconsistencies caused by loop closure. To address these issues, this paper presents PLCS-SLAM, a visual localization and mapping system for dynamic and low-texture environments. The system combines depth-aware semantic gating with probabilistic geometric verification to suppress dynamic outliers while preserving valid static features. A tightly coupled point-line optimization framework is introduced, in which continuous motion confidence is propagated from the front-end to the back-end information matrix and line features are used to strengthen structural constraints. A keyframe-anchored semantic octree mapping module is further developed to reduce structural ghosting and maintain map consistency during pose updates. Experiments on the TUM RGB-D benchmark, KITTI, and a real-world UAV dataset show lower trajectory errors than several evaluated baselines, while paired map evaluations indicate reduced local structural misalignment on KITTI.