Dense SLAM Using 3-D Gaussian Splatting With IMU and Geometric Constraints

Simultaneous Localization and Mapping (SLAM) is a foundational technology for autonomous systems, augmented reality, and robotics, enabling real-time localization and environment perception. Recently, Gaussian Splatting has been widely adopted in SLAM frameworks due to its capabilities for explicit geometric representation, real-time photorealistic rendering, and efficient differentiable optimization. However, although existing monocular methods exploit line features and IMUs to improve Gaussian reconstruction, errors introduced during IMU initialization and the lack of depth supervision during optimization often lead to uncertain artifacts, limiting overall system performance. To address these limitations, we propose a tightly-coupled monocular-inertial framework that leverages high-frequency IMU observations to enhance both localization and Gaussian mapping. Our approach incorporates an efficient IMU initialization scheme and utilizes both point-line features and IMU observations for robust tracking. Moreover, IMU-supervised depth estimation and normal maps provide metric-scale geometric constraints, which are further combined with structural lines to generate planar hypotheses, densifying Gaussian primitives in low-texture regions. A pyramid-based Gaussian optimization framework is then employed, where multi-dimensional losses drive the progressive refinement of both global geometric structures and local texture details. Extensive evaluations on the Replica, TUM RGB-D, ScanNet, OpenLORIS, and KITTI datasets demonstrate that our method achieves more stable tracking and higher-quality reconstructions than state-of-the-art approaches, while maintaining real-time performance.