Abstract
A dynamic visual SLAM system named XYG-SLAM is proposed in this paper, which integrates semantic and geometric constraints to address trajectory drift and mapping ‘ghosting’ artifacts in traditional SLAM systems caused by moving object interference in dynamic scenes. Precise state awareness of dynamic objects is achieved through the integration of instance segmentation with multi-view geometric constraints. The feature culling strategy is adaptively adjusted based on the dynamic level of the scene, thereby maximizing feature utilization while ensuring robustness. To further enhance matching robustness, the deep learning-based XFeat feature is incorporated. Furthermore, a dynamic foreground segmentation method based on inverse depth Gaussian modeling is introduced, which effectively strips away dynamic objects and enables the construction of high-quality static maps. Experimental results on the TUM RGB-D and self-collected datasets demonstrate that XYG-SLAM achieves better overall trajectory stability than representative dynamic SLAM methods, especially in highly dynamic sequences. Static map construction results further show that XYG-SLAM effectively preserves edge structures in static maps and suppresses dynamic ghosting artifacts, providing a reliable solution for robust visual SLAM in complex dynamic environments.

This publication has 30 references indexed in Scilit: