Semantically Constrained SLAM Using Reliable Object Landmarks in Dynamic Vehicle Environments

Semantic information has been increasingly incorporated as pose optimization constraints into visual simultaneous localization and mapping (vSLAM) for autonomous vehicles against complex environments. However, existing semantically constrained SLAM systems are still vulnerable to incomplete object observations and mobile objects, both of which are prevalent in dynamic vehicle environments. To mitigate these challenges, we propose a semantically constrained SLAM method by extracting and utilizing reliable object landmarks. This method first detects the motion status of objects using both geometric and semantic information, specifically tailored to address degenerate motions in driving scenarios. Subsequently, object-level landmarks are constructed from the identified static features and objects. These landmarks are then used to establish semantic constraints, which are adaptively weighted based on the reliability of each landmark, for camera pose optimization. Experimental results using a self-collected vehicle-borne dataset characterized by multiple categories of dynamic objects and severe partial observations demonstrate that the proposed system improves the 3D positioning accuracy by approximately 68% and 29% compared to two state-of-the-art semantic vSLAM systems. Experimental evaluations on diverse sequences from the public KITTI and TUM RGB-D datasets are also performed to further validate the superior positioning accuracy and robustness provided by the proposed system across both dynamic outdoor and indoor environments.