DSG-SLAM: a visual SLAM for removing dynamic features and dense mapping in dynamic environments

Abstract

Abstract Visual SLAM, widely applied in autonomous driving and robotic navigation, enables real-time environmental perception and high-precision mapping through sensors, significantly reducing reliance on pre-known environmental information. However, it faces critical challenges in dynamic environments, where dynamic objects interfere with feature point matching, leading to degraded localization accuracy, increased trajectory errors, and insufficient map density. To address these issues, this paper proposes DSG-SLAM, a Dynamic-Semantic-Gaussian SLAM system based on ORB-SLAM3, designed to resolve trajectory errors caused by dynamic objects and generate denser maps in dynamic environments. DSG-SLAM processes images from RGB-D sensors to extract ORB feature points. It integrates dynamic object detection through Segformer and a depth map-based dynamic object detection, enabling precise removal of ORB features in dynamic regions to enhance localization accuracy. Additionally, the LaMa inpainting network reconstructs static backgrounds by repairing dynamic regions, followed by high-quality dense mapping via 3D Gaussian Splatting. Evaluated on the TUM and Bonn dataset, DSG-SLAM achieves 97.47% reduction in absolute trajectory error and 97.62% reduction in standard deviation compared to ORB-SLAM3 in high-dynamic scenarios. It also outperforms state-of-the-art dynamic SLAM methods in localization precision and robustness. DSG-SLAM substantially enhances both localization precision and map density in dynamic environments, providing a more reliable solution for robotic applications in such scenarios.