TriGS-SLAM: Tristate Visibility Modeling for Robust 3DGS-SLAM in Dynamic Scenes Based on RGB-D Sensor

Abstract

Simultaneous localization and mapping (SLAM) systems based on 3-D Gaussian splatting (3DGS) have garnered widespread attention due to their real-time high-fidelity rendering capabilities. However, in dynamic scenes, the interpretation of Gaussian visibility at the boundaries of dynamic occlusions is still unstable, often manifesting as boundary flickering, ghosting, residual artifacts, and geometric interpretation errors, which further interfere with pose estimation and map updates. The reason is that dynamic occlusion is not a simple binary switching process, but a continuous transition from fully visible to completely invisible. Traditional binary mask processing causes structural errors in information utilization for continuous occlusion changes, whereas relying solely on continuous opacity to handle discrete visibility switching can easily lead to boundary flickering and trailing effects. To this end, we propose a three-state visibility modeling method driven by multisource evidence for dynamic scenes captured by red-green-blue-depth (RGB-D) sensors. This method explicitly introduces a partial visibility intermediate state between the traditional binary states of visible and invisible, thereby transforming dynamic occlusion handling from a traditional post hoc residual-correction strategy into a visibility interpretation mechanism oriented toward the generative process. On this basis, we further use the three-state posterior as a soft constraint weight to uniformly regulate rendering, pose optimization, and mapping, enabling the system to form more stable and consistent observation interpretations in dynamic boundaries and occluded areas. Experimental results show that the proposed method can effectively alleviate boundary flickering and occlusion artifacts, and outperform existing state-of-the-art methods in terms of pose tracking accuracy and map reconstruction quality on real-world dynamic-scene datasets.