Visual SLAM Algorithm Based on Instance Segmentation in Dynamic Scene

To address the decline in localization accuracy caused by the static environment assumption in traditional visual Simultaneous Localization and Mapping (SLAM) systems under dynamic scenes, this paper proposes a robust dynamic feature point removal method by integrating YOLOv8 instance segmentation with ORB-SLAM3. The system introduces a dynamic perception thread into the ORB-SLAM3 framework, leveraging the YOLOv8-seg network to achieve instance-level dynamic object detection. Combined with optical flow tracking residuals and multi-view epipolar geometric constraints, a probabilistic discrimination model for dynamic feature points is constructed to enable pixel-level refinement of dynamic interference. During dynamic object detection, YOLOv8-seg distinguishes between different instances of the same category compared to traditional semantic segmentation networks, thereby reducing static feature loss caused by indiscriminate removal of semantic regions. In pose estimation, dynamic probability weighting optimizes feature point matching weights, preserving usable static features in partially occluded or semi-static scenarios. Evaluations on the TUM dynamic dataset demonstrate that the proposed system reduces the Absolute Trajectory Error (ATE) by approximately 41.2% compared to the original ORB-SLAM3 in high-speed motion and densely dynamic scenarios. While maintaining real-time performance, it significantly outperforms DynaSLAM and DS-SLAM in discriminating partially dynamic objects, providing a highly robust solution for SLAM deployment in dynamic environments.