CSM-SLAM: A Dynamic Visual SLAM Approach Based on Cross-Frame Semantic Matching

In indoor dynamic environments, robots rely on visual SLAM to perceive and understand their surroundings. However, the presence of moving objects violates the static-world assumption, thereby degrading the accuracy and robustness of visual SLAM localization. Although most dynamic visual SLAM approaches achieve satisfactory localization performance in dynamic scenarios, they typically reject dynamic points solely based on semantic information, without considering semantic associations across frames. This limitation results in insufficient exploitation of temporal semantic consistency, which could otherwise enhance robustness in dynamic environments. To address this issue, we propose CSM-SLAM, a dynamic visual SLAM system that incorporates a cross-frame semantic matching (CSM) strategy. By leveraging semantic patches and their temporal correspondences, our approach enables robust identification and removal of dynamic feature points, thereby refining pose estimation. CSM-SLAM is developed on top of ORB-SLAM3 and evaluated on the TUM and BONN datasets. Experimental results demonstrate that our method can accurately eliminate dynamic feature points, achieving up to a 90% improvement in localization accuracy compared to ORB-SLAM3 in dynamic scenarios. The code for the cross-frame semantic matching method proposed in this paper is available at https://github.com/wyyaa123/CSM.git.