DySeG-SLAM: A Dynamic SLAM System Integrating YOLACT and Multi-View Geometry

To address feature mismatching and localization degradation caused by dynamic objects in visual SLAM systems operating under dynamic environments, this paper proposes DySeG-SLAM—a dynamic SLAM framework integrating semantic-geometric constraints. The system enhances ORB-SLAM2 through four primary modifications. First, a Channel-Spatial Dual Attention Mechanism (CSAM) is integrated into the YOLACT instance segmentation network to improve small object detection. Second, a pyramid iterative Lucas-Kanade (LK) optical flow method is implemented for multi-scale feature tracking in static regions. Third, a multi-view geometric approach combining disparity angle analysis and depth projection error evaluation is introduced to identify dynamic objects parallel to camera motion. Finally, a dynamic probability model incorporating historical states, epipolar constraints, and multi-view geometry is established to stabilize detection consistency against single-frame errors. Evaluations on the TUM dataset demonstrate that DySeG-SLAM achieves superior localization accuracy compared to ORB-SLAM2 in dynamic scenarios, with translation and rotation errors reduced by over 92%. The results indicate enhanced robustness against dynamic feature interference while maintaining computational efficiency.