A Mobile Robot SLAM Algorithm Based on Semantic Dynamic Feature Elimination

To mitigate the degradation of localization accuracy in mobile robots caused by dynamic object interference in dynamic environments, this paper proposes a visual SLAM method based on semantic dynamic feature elimination. First, the YOLOv11n-seg architecture is optimized by incorporating the Scale Sequence Fusion (ScalSeq), Triple Feature Encoder (Zoom_cat), and Channel and Position Attention Mechanism from ASF-YOLO, together with the Semantics and Detail Infusion (SDI) module from U-Net V2, thereby improving the boundary precision and multi-scale adaptability of dynamic object masks. Second, within the ORB-SLAM2 framework, the lightweight integration of the YOLO semantic segmentation model into the tracking thread enables accurate removal of feature points belonging to dynamic objects during the feature extraction stage. Finally, a dynamicity evaluation and semantic label injection mechanism for map points, based on multi-frame observations, is introduced into the local mapping thread; during backend optimization, the constraint weights of dynamic map points are adaptively suppressed according to their semantic labels. Experimental results on public datasets show that the improved model achieves a 1.70% increase in mean average precision (mAP). Compared with the original ORB-SLAM2, the improved algorithm substantially reduces the absolute trajectory error, with the root mean square error (RMSE) decreasing by up to 45.27% and the mean error (Mean) by up to 47.84%. In the real-robot experiments, the number of dynamic feature points extracted by the improved algorithm is significantly reduced compared with that extracted by the original algorithm. The proposed method can therefore effectively handle various indoor and outdoor dynamic scenarios, significantly enhancing the localization accuracy and robustness of mobile robots in dynamic environments.