Enhanced Visual SLAM Accuracy in Dynamic Environments via YOLOv5-Based Dynamic Feature Filtering

To address the challenges faced by visual SLAM algorithms in dynamic environments, where moving objects can interfere and degrade localization and mapping accuracy, this paper introduces a dynamic object detection thread to the ORBSLAM3 system. This enhancement improves the adaptability of the SLAM system to environmental changes and its practicality for mapping tasks. The dynamic object detection thread employs the YOLOv5 object detection algorithm to identify dynamic objects in the scene. Corresponding feature points associated with these objects are subsequently removed to mitigate their impact, thereby improving localization and mapping accuracy. Additionally, a point cloud mapping thread is designed to generate a dense point cloud map using keyframes and their associated poses. This approach not only eliminates ghosting effects caused by dynamic objects but also retains the static feature points in the map, enabling precise dense map reconstruction. Experiments conducted on the TUM dataset demonstrate that, compared to the original ORB-SLAM3 algorithm, the improved method achieves an enhancement of 96.9 % in absolute trajectory accuracy and 92.8 % in relative pose accuracy in highly dynamic sequences. These results indicate that the proposed algorithm effectively removes dynamic feature points and enhances the localization accuracy and pose estimation precision of visual SLAM in dynamic environments.