Dynamic visual SLAM algorithm combining object detection and point-line fusion

To address the issue of damaged accuracy and robustness in Simultaneous Localization and Mapping (SLAM) algorithms caused by moving objects in the scene, a visual SLAM algorithm based on lightweight object detection and point-line fusion is proposed. Firstly, the ORB-SLAM3 is enhanced by integrating the Line Segment Detector (LSD) algorithm for feature line detection and the Line Band Descriptor (LBD) algorithm for feature line description, allowing for the simultaneous extraction and fusion of feature points and lines. Then, the lightweight YOLOv8n-seg model is employed to detect dynamic objects in the environment and obtain their bounding boxes. The dynamic feature points and lines detected by YOLOv8n-seg are removed from the improved ORB-SLAM3's feature set, and the remaining static feature points and lines are used for weighted pose estimation. Finally, a dense point cloud map of the static environment is constructed. Experimental results show that compared to ORB-SLAM3 and YOLO-ORB-SLAM3, the proposed algorithm reduces the RMSE in ATE by 63.71% and 24.19%, respectively, in dynamic scenes, effectively improving the accuracy of SLAM algorithms in dynamic scenarios. This has significant implications for enhancing the navigation and environmental perception capabilities of autonomous robots and equipment in industrial automation systems, providing more accurate and robust solutions for the advancement of intelligence and automation in manufacturing processes.