A Monocular SLAM System for Large-Scale 3D Reconstruction

In this paper, a monocular SLAM system for large-scale 3D reconstruction is proposed, using 3D Gaussian Splatting as the underlying map representation. By cleverly integrating lightweight depth estimation, sparse feature tracking, and Gaussian optimization strategies, this system overcomes the reliance on depth sensors and the high computational demands associated with traditional techniques. Incorporating a monocular depth estimation model, our system takes monocular RGB images as input to achieve high-precision reconstruction without the need for depth sensors. Furthermore, through the use of effective Gaussian models and selective elimination techniques, it actively incorporates and fine-tunes Gaussians, leading to a notable decrease in computational demands. Validation of the system has been carried out experimentally across multiple public datasets and an in-house developed virtual Mars simulation, revealing remarkable performance in tracking fidelity, memory footprint, and real-time processing capabilities. This provides an effective solution for future large-scale scene reconstruction tasks.