PDGS-SLAM: Photorealistic Dense Gaussian Splatting SLAM for Dynamic Scenes

Abstract

3D Gaussian Splatting has enabled photorealistic dense mapping for visual Simultaneous Localization and Mapping (SLAM). Nevertheless, most existing pipelines are designed for static scenes and thus degrade under dynamic occlusions, leading to unstable tracking and noticeable mapping artifacts. Prior dynamic solutions often depend on strong semantic priors or frame-wise uncertainty prediction, which are prone to temporal inconsistency and confusion in highly textured regions. To address these issues, we present PDGS-SLAM, a robust dense SLAM system for dynamic environments. The framework introduces two key designs: first, a Temporally Consistent Occlusion-aware Module that combines depth-guided warping with geometric constraints to enforce temporally coherent occlusion estimation, suppressing inter-frame noise without requiring semantic labels; second, a Dynamic Frequency Enhancement (DFE) strategy that applies multi-scale frequency regularization to preserve high-frequency static structures during occlusion modeling, thereby improving reconstruction fidelity in texture-rich areas. In addition, we adopt an occlusion-guided coarse-to-fine tracking scheme to reduce pose drift in dynamic scenes. Extensive experiments on public benchmarks show that PDGS-SLAM achieves state-of-the-art tracking accuracy and improved rendering quality, with consistent PSNR gains while maintaining competitive SSIM and LPIPS performance.