GLC-SLAM: Gaussian Splatting SLAM with Efficient Loop Closure (2409.10982v1)

Abstract: 3D Gaussian Splatting (3DGS) has gained significant attention for its application in dense Simultaneous Localization and Mapping (SLAM), enabling real-time rendering and high-fidelity mapping. However, existing 3DGS-based SLAM methods often suffer from accumulated tracking errors and map drift, particularly in large-scale environments. To address these issues, we introduce GLC-SLAM, a Gaussian Splatting SLAM system that integrates global optimization of camera poses and scene models. Our approach employs frame-to-model tracking and triggers hierarchical loop closure using a global-to-local strategy to minimize drift accumulation. By dividing the scene into 3D Gaussian submaps, we facilitate efficient map updates following loop corrections in large scenes. Additionally, our uncertainty-minimized keyframe selection strategy prioritizes keyframes observing more valuable 3D Gaussians to enhance submap optimization. Experimental results on various datasets demonstrate that GLC-SLAM achieves superior or competitive tracking and mapping performance compared to state-of-the-art dense RGB-D SLAM systems.

• The paper introduces GLC-SLAM, a SLAM system that leverages 3D Gaussian splatting and a hierarchical loop closure mechanism to ensure global consistency in large-scale environments.
• The method employs an uncertainty-minimized keyframe selection and frame-to-model tracking to significantly reduce map drift and enhance tracking performance.
• Comprehensive evaluations demonstrate that GLC-SLAM outperforms state-of-the-art dense RGB-D SLAM methods in metrics like ATE, PSNR, and SSIM for both tracking and mapping.

Overview of GLC-SLAM: Gaussian Splatting SLAM with Efficient Loop Closure

The paper "GLC-SLAM: Gaussian Splatting SLAM with Efficient Loop Closure" presents a novel approach in the domain of Simultaneous Localization and Mapping (SLAM) that addresses several prevalent challenges in existing methods, particularly those employing 3D Gaussian Splatting (3DGS) representations. SLAM systems are integral to applications requiring accurate scene reconstruction and real-time camera tracking, such as in robotics and augmented reality.

Key Contributions

The main contributions of the paper are as follows:

1. Introduction of GLC-SLAM: The researchers propose GLC-SLAM, a system that leverages 3D Gaussian Splatting for robust frame-to-model tracking and guarantees global consistency in large-scale environments through hierarchical loop closure.
2. Efficient Loop Closure Module: The system incorporates a multi-layered loop closure strategy, which includes global-to-local loop detection, pose graph optimization, and direct updates to the map, significantly reducing accumulated errors and map drift.
3. Uncertainty-Minimized Keyframe Selection: The authors introduce a method that enhances the selection of keyframes by prioritizing those observing more valuable and stable 3D Gaussians, thus improving submap optimization.

Technical Approach

Scene Representation

GLC-SLAM uses 3D Gaussian distributions for scene representation. Each submap within the system is composed of multiple 3D Gaussians parameterized by mean, covariance, opacity, and color values. This representation allows efficient rendering and real-time performance while maintaining high reconstruction fidelity.

Tracking

The system employs a frame-to-model tracking approach, initializing the camera pose based on the previous frame and refining it by minimizing a tracking loss function. This reduces gross errors and ensures consistent tracking even in dynamically changing environments.

Local Mapping

Local mapping in GLC-SLAM is achieved by incrementally building 3D Gaussian submaps and anchoring them to global keyframes. The uncertainty-minimized keyframe selection strategy further refines this process by selecting frames that contribute most effectively to the optimization.

Loop Closing

The loop closing module is the cornerstone of the system’s ability to handle map drift and accumulated errors. Hierarchical loop closure is conducted in two stages:

1. Global loop detection when new submaps are created.
2. Local loop detection during the continuous mapping process.

Pose graph optimization then aligns keyframe poses to correct the trajectory and reduce errors by updating the associated 3D Gaussian map.

Evaluation and Results

The paper provides comprehensive evaluations of GLC-SLAM using multiple datasets, including Replica, TUM-RGBD, and ScanNet. The results show that GLC-SLAM outperforms existing state-of-the-art dense RGB-D SLAM methods in terms of:

1. Tracking Accuracy: The system demonstrates superior tracking accuracy as reflected in the Average Trajectory Error (ATE) metrics across varied datasets and scenes.
2. Rendering Quality: GLC-SLAM achieves higher scores in PSNR, SSIM, and LPIPS, indicating better visual fidelity of the reconstructed scenes.
3. Mapping Performance: Quantitative metrics like Depth L1 and F1-score showcase the accuracy and completeness of the reconstructed maps.

The system also achieves competitive runtime and memory performance, making it feasible for real-time applications.

Implications and Future Work

The introduction of GLC-SLAM represents a significant advancement in the domain of dense RGB-D SLAM systems. The hierarchical loop closure mechanism and uncertainty-minimized keyframe selection improve the robustness and accuracy of 3D mapping in large-scale environments. Future developments could focus on:

1. Further Optimization: Enhancing the efficiency of the loop closure process and minimizing computational overhead.
2. Extended Real-World Applications: Applying the approach to more diverse environmental conditions and incorporating more complex dynamic elements.
3. Integration with Other Sensors: Extending the system to integrate other sensor modalities, such as LIDAR, which may provide additional constraints and improve mapping accuracy in challenging scenarios.

Conclusion

The GLC-SLAM system addresses key challenges in 3DGS-based SLAM methods by introducing effective loop closure mechanisms and improving keyframe selection strategies. The results from extensive evaluations demonstrate its effectiveness in achieving robust, accurate, and real-time SLAM, making it a substantial contribution to the field. Further research and practical implementations will likely continue to expand the applications and capabilities of such systems in the ever-evolving landscape of autonomous systems and immersive technologies.