A Visual Slam Algorithm for Dynamic Environments Based on Instance Segmentation and Dynamic Probability Estimation

Traditional visual SLAM systems often suffer from degraded localization accuracy and feature mismatches in highly dynamic indoor environments due to moving object interference. To address this challenge, we propose a semantic-enhanced SLAM algorithm based on the ORB-SLAM3 framework. Our system integrates the YOLOv11-seg instance segmentation network for precise dynamic object detection and introduces an adaptive mask dilation strategy combined with multi-scale edge enhancement to refine dynamic region boundaries. Furthermore, we design a dynamic probability estimation module that leverages depth-based clustering and conditional random fields to identify and remove residual features on non-prior dynamic objects. Only reliable static feature points are used for pose estimation, improving matching stability and trajectory accuracy. Experimental results on the TUM RGB-D dataset and real-world indoor scenarios demonstrate that the proposed method significantly improves localization accuracy in highly dynamic environments while ensuring processing speed suitable for online SLAM applications.