A Dynamic Visual SLAM Method Based on Semantic and Scene Flow Constraints

Abstract

Traditional visual SLAM methods generally rely on the static-scene assumption and may suffer substantial localization degradation in the presence of moving objects. To address this issue, this paper proposes a dynamic visual SLAM method based on semantic and scene-flow constraints and integrates it into the ORB-SLAM3 framework. First, YOLO11s is used to detect potentially dynamic objects and construct semantic prior regions. Second, adjacent-frame depth information and relative camera motion are combined to build a dynamic-feature discrimination model based on depth-normalized scene-flow residuals and deviations from a global motion reference. Finally, a dual-threshold strategy applies different criteria inside and outside the semantic prior regions, retaining only static features for pose estimation and back-end optimization. Experimental results show that, in terms of ATE RMSE, the proposed method reduces the error by an average of 69.20% on the TUM indoor dynamic sequences and 17.87% on the KITTI outdoor road sequences relative to ORB-SLAM3, while also achieving lower errors than representative dynamic SLAM methods on the evaluated sequences.