Improvement Scheme of VSLAM System in Dynamic Map Point Removal

Precision localization and mapping within indoor environments represent a significant challenge for Visual Simultaneous Localization and Mapping (VSLAM) systems. While existing methodologies, such as ORB-SLAM3, demonstrate high accuracy in static settings, their capacity for precise localization and robustness is substantially compromised in the presence of dynamic objects. Consequently, mitigating the impact of dynamic objects on the localization and mapping processes of VSLAM systems is an imperative issue that demands immediate resolution. To address this challenge, this paper introduces an enhanced ORB-SLAM3 approach that integrates deep learning with dynamic environment mitigation, tailored specifically for indoor settings characterized by motion. By amalgamating the state-of-the-art YOLOv10 object detection algorithm with the Lucas-Kanade (LK) optical flow technique to eliminate dynamic map points, the proposed algorithm effectively neutralizes the influence of dynamic objects on pose estimation, thereby bolstering the robustness of VSLAM systems in the context of dynamic scene localization. Empirical findings indicate that, in comparison to the original ORB-SLAM3, the refined algorithm has realized a marked enhancement of over 95% in Average Trajectory Error (ATE) within highly dynamic environments, with improvements in Relative Trajectory Error (RTE) also exceeding 95%, without compromising the system's operational real-time integrity.