SGU-SLAM: Semantic-Geometric Fusion and Uncertainty Embedding for Dynamic SLAM

Reliable localization and mapping in highly dynamic environments are key challenges for Visual SLAM due to the violation of the static world assumption. Deep learning-based semantic VSLAM methods face a trade-off: instance segmentation is computationally costly, while object detection's coarse bounding boxes discard valid background features, degrading pose accuracy. To address this, we propose a robust RGB-D SLAM framework based on ORB-SLAM3, integrating a lightweight object detector with a probabilistic semantic-geometric fusion mechanism. An Adaptive Gaussian Mixture Model recovers static background features within dynamic bounding boxes, avoiding heavy segmentation networks. Additionally, a geometry-based probabilistic filter identifies unknown dynamic objects via epipolar constraints. Instead of hard dynamic feature elimination, we incorporate temporal belief states into an uncertainty-embedded factor graph optimization to dynamically down-weight unreliable landmarks. Extensive experiments on the TUM RGB-D dataset show our system outperforms state-of-the-art dynamic SLAM methods in high-dynamic sequences while maintaining real-time performance.