EGC-SLAM: enhanced geometric constraints for monocular Gaussian SLAM

Using only monocular sensors for photorealistic scene reconstruction has faced significant challenges. Recent 3-D Gaussian splatting (3DGS)-based simultaneous localization and mapping (SLAM) systems have demonstrated remarkable progress, enabling dense and photorealistic scene reconstruction. However, existing methods still exhibit limitations in geometric accuracy and heavily rely on depth input. To address these issues, this article proposes a monocular Gaussian SLAM framework that organically integrates an uncertainty-aware tracking mechanism with a geometry-consistent Gaussian mapping method. Specifically, the system first estimates initial camera poses by matching point-cloud maps generated by the MASt3R network. To enhance robustness in dynamic environments, we introduce an uncertainty-aware mechanism into the pose estimation module. For mapping, we design a multimodal geometry-aware point-cloud distribution model to reduce representational redundancy during Gaussian initialization. Meanwhile, we propose a spatially varying depth (S-V Depth) rendering mechanism under the local projection model of 3DGS, which effectively improves geometric consistency and alleviates depth distortions caused by overlapping surfaces. Experimental results demonstrate that the proposed method not only maintains real-time performance but also significantly enhances rendering quality and reconstruction accuracy, offering a new direction for the development of monocular Gaussian SLAM systems.