Dynamic Visual SLAM Based on Segment-Everything and Instance-Level Temporal Inference

Simultaneous localization and mapping (SLAM) is a fundamental capability for autonomous localization and environmental perception in mobile robots. Conventional RGB-D visual SLAM generally relies on a quasi-static scene assumption. In open dynamic environments, independently moving objects may disrupt inter-frame data association and violate static rigid-body geometric consistency. These effects can lead to pose drift, tracking degradation, and map corruption. To address this problem, this paper proposes Dynamic Instance Filtering SLAM (DIF-SLAM), an RGB-D visual SLAM framework for unknown dynamic scenes based on open-world instance segmentation and instance-level temporal inference. The proposed method first normalizes open-world instance segmentation outputs into a mutually exclusive single-channel label map. It then establishes instance-level inter-frame association and constructs velocity observations from 3D static-compensation residuals. A two-state hidden Markov model with dual-channel posterior enhancement is further used for temporally stable motion-state inference. The inferred instance states are projected to pixel-level dynamic masks for dynamic observation suppression in Tracking and Dense Mapping. Experiments on the TUM RGB-D dataset, the Bonn RGB-D dataset, and real-world scenes show that DIF-SLAM improves localization accuracy and reconstruction quality in sequences with strong dynamic rigid-body interference. Its gains are more limited in low-dynamic scenes and non-rigid motion scenarios. Overall, the proposed method improves the temporal stability of instance-level motion-state estimation under unknown dynamic interference and reduces the impact of dynamic observations on pose estimation and map reconstruction.