DMS-SLAM: semantic visual SLAM based on deep mask segmentation in dynamic environments

Abstract Visual simultaneous localization and mapping (VSLAM), which is a core technology for intelligent mobile robots, is typically based on static environment assumptions. However, it encounters challenges in dynamic scenes such as accurately distinguishing between static and dynamic feature points and high computational costs. These issues lead to significant deviations in camera tracking and introduce ghosting in mapping. To address these issues, we propose DMS-SLAM, a fast semantic visual SLAM based on deep mask segmentation. First, semantic segmentation is performed using the YOLOv8s-seg network, followed by a mask correction algorithm based on depth consistency and a missed detection compensation algorithm based on relative displacement invariance, ensuring precise segmentation of dynamic objects within the semantic thread. Then, the segmented objects are classified according to dynamic probability, and the motion of potential dynamic objects is determined through a depth mask association algorithm. Dynamic feature points are eliminated by utilizing epipolar geometric constraints to improve localization accuracy. Finally, local point clouds with dynamic objects removed are stitched together to construct a global dense point cloud map, presenting a static scene free of ghosting. Experiments on the Technical University of Munich RGB-D dataset and in real-world scenarios show that the method enhances absolute trajectory accuracy by an average of 96.19% and relative pose accuracy by 94.27% over the ORB-SLAM2 system in highly dynamic environments. Compared to other dynamic semantic visual SLAM algorithms, DMS-SLAM improves operational efficiency by over 51.78%, significantly reducing runtime and generating clear and accurate global 3D point cloud maps, which fully demonstrates its superiority in terms of accuracy, stability and efficiency.