Enhancing ORB-SLAM3 Pose Estimation in Dynamic Scenes with YOLOv5 Object Detection

This research paper introduces an improved ORB-SLAM3, which is integrated with YOLO to achieve real-time dynamic feature point elimination and pose optimization. ORB-SLAM3 uses the ORB feature extractor to extract corner points, uses descriptors to match supporting points, and finally estimates the camera pose. However, in some dynamic scenes, the motion of dynamic objects will affect the pose estimation accuracy of SLAM, resulting in incorrect camera pose estimation. In order to solve this problem, it is proposed to add YOLO into ORB-SLAM3, and use YOLO’s target detection to detect prior dynamic objects in image frames and eliminate feature points in dynamic objects, so as to reduce the impact of dynamic objects on camera pose estimation. The idea was evaluated on a publicly available dataset, which confirmed that pose estimation in dynamic scenarios was more accurate than the benchmark ORB-SLAM3.