Abstract:

The original ORB descriptor algorithm has a low matching accuracy and long matching time, the positioning accuracy and robustness of the SLAM(simultaneous localization and mapping) system are severely disturbed by moving objects in dynamic scenes, and the ORB-SLAM3 system is incapable of constructing dense maps. To address the above problems, this paper proposes an improved ORB-SLAM3 based on the BEBLID descriptor and object detection. A lightweight YOLOv5s dynamic object detection network and dynamic feature removal module are fused with the tracking thread to improve the system's positioning accuracy. Replacing the original feature description algorithm, an improved local image descriptor BEBLID with higher efficacy is combined with the original ORB feature extraction method to enhance the expressiveness and description efficiency of images, ensuring a more accurate and efficient feature matching. A dense mapping thread is added to construct dense point cloud maps based on keyframes and correspondingposes. Experiments on the publicly available TUM RGB-D dataset show that compared with the original ORB-SLAM3, the proposed algorithm has a 7% higher feature matching accuracy; the system's positioning accuracy is improved by more than 98% in high dynamic environments and by up to 60% in low dynamic environments, showing a better positioning function in dynamic environments; a three-dimensional dense point cloud map is constructed, laying a foundation for future applications in robot autonomous navigation, obstacle avoidance, and path planning.

Key words: simultaneous localization and mapping (SLAM), ORB-SLAM3, BEBLID, YOLOv5s, dense mapping