A Visual SLAM Approach for Unmanned Systems: Dynamic Feature Point Detection and Elimination by Combining YOLOv5 and Epipolar Constraints

In the field of computer vision, particularly in visual Simultaneous Localization and Mapping (SLAM) systems, environments are typically assumed to be static, with minimal lighting changes and no object deformation. However, real-world environments are far more complex, filled with dynamic elements such as moving people, pets, furniture indoors, and cars and pedestrians outdoors. These dynamic elements introduce additional challenges to visual SLAM systems, as feature points may fall on moving objects, interfering with pose estimation, increasing computational errors, causing tracking failures, and reducing the accuracy of loop detection, ultimately affecting the system's overall precision and robustness. To address these challenges, this paper proposes a novel strategy for dynamic feature point detection and elimination by combining YOLOv5 and epipolar constraints. YOLOv5's efficient object detection capabilities accurately identify dynamic objects in images, while epipolar constraints provide a geometric method to evaluate the consistency of feature points across consecutive frames. This effectively identifies and eliminates feature points on dynamic objects. The proposed method not only improves the accuracy and stability of target detection in dynamic environments but also provides more reliable feature information for pose estimation and map construction in visual SLAM systems, enhancing overall system performance and adaptability. The algorithm was tested on the TUM dataset to validate its optimization and computational effectiveness. Experimental results demonstrate that our approach significantly improves localization accuracy and robustness in complex dynamic scenes.