Mobile educational robots require robust visual SLAM to operate safely in dynamic indoor environments. However, conventional systems such as ORB-SLAM3 assume static scenes and degrade under human motion due to excessive removal of static features (e.g., ATE = 0.259 m on TUM RGB-D D \(\:{fr3}_{{w}_{xyz}}\)). This paper proposes a unified dynamic-robust RGB-D SLAM framework tailored for resource-constrained platforms. An enhanced YOLOv8s detector incorporating Partial Convolution within the Faster_C2f module achieves real-time dynamic object detection at 54.6 fps on CPU (3.7 × faster than baseline). A two-stage static feature preservation strategy combines DBSCAN-based depth clustering for semantic mask refinement with epipolar-geometry-based motion classification, recovering up to 81% of static features in highly dynamic scenes. Additionally, an adaptive Point–Line Fusion approach based on Plücker coordinates improves robustness in low-texture regions. The complete system runs at 14.6 fps on CPU. Experiments on six TUM RGB-D sequences show reductions of 81.28% in Absolute Trajectory Error and 93.59% in Relative Pose Error compared to ORB-SLAM3 (p < 0.001), while achieving over 13× higher frame rates than DynaSLAM with comparable reconstruction quality.
