RED-SLAM: real-time and effective RGB-D SLAM with spatial-geometric observations and fast semantic perception for dynamic environments

Abstract Most visual simultaneous localization and mapping (vSLAM) methods assume a static scene, limiting their effectiveness in complex, real-world dynamic environments. This paper presents RED-SLAM-a real-time dynamic visual SLAM method based on the ORB-SLAM3 framework for RGB-D sensors, designed to effectively address the impact of dynamic objects. RED-SLAM leverages spatial-geometric observations combined with semantic cues to identify dynamic points within the field of view, thereby utilizing only static points for state estimation. In the geometric verification module, the initial distinction between static and dynamic points is achieved by checking the spatial projection ray distance error based on the matching between map points and feature points. To conserve computational resources, semantic segmentation is performed exclusively on designated semantic frames, which are constructed based on the changes of dynamic objects. The detected dynamic semantic objects are subsequently spread to successive frames using a semantic propagation technique. All map points associated with dynamic semantic objects are excluded to further enhance the identification accuracy of dynamic map point. Compared to existing methods that apply semantic segmentation across all frames or keyframes, RED-SLAM only performs segmentation when dynamic objects change, improving system’s performance and efficiency. Experimental results on public datasets and real-world scenes demonstrate that our method enhances pose estimation accuracy in dynamic environments, achieving competitive accuracy compared to state-of-the-art methods, while maintaining reliable real-time performance.