DGS-SLAM: Gaussian Splatting SLAM in Dynamic Environment (2411.10722v1)

Abstract: We introduce Dynamic Gaussian Splatting SLAM (DGS-SLAM), the first dynamic SLAM framework built on the foundation of Gaussian Splatting. While recent advancements in dense SLAM have leveraged Gaussian Splatting to enhance scene representation, most approaches assume a static environment, making them vulnerable to photometric and geometric inconsistencies caused by dynamic objects. To address these challenges, we integrate Gaussian Splatting SLAM with a robust filtering process to handle dynamic objects throughout the entire pipeline, including Gaussian insertion and keyframe selection. Within this framework, to further improve the accuracy of dynamic object removal, we introduce a robust mask generation method that enforces photometric consistency across keyframes, reducing noise from inaccurate segmentation and artifacts such as shadows. Additionally, we propose the loop-aware window selection mechanism, which utilizes unique keyframe IDs of 3D Gaussians to detect loops between the current and past frames, facilitating joint optimization of the current camera poses and the Gaussian map. DGS-SLAM achieves state-of-the-art performance in both camera tracking and novel view synthesis on various dynamic SLAM benchmarks, proving its effectiveness in handling real-world dynamic scenes.

• The paper introduces a novel SLAM framework integrating Gaussian Splatting with robust mask generation and loop-aware optimization.
• It achieves state-of-the-art results in camera tracking and mapping, demonstrating improved ATE, PSNR, SSIM, and LPIPS metrics in dynamic settings.
• The approach effectively filters dynamic elements, broadening SLAM applications in robotics, autonomous navigation, and AR/VR environments.

Overview of DGS-SLAM: Gaussian Splatting SLAM in Dynamic Environments

In this work, the authors present Dynamic Gaussian Splatting SLAM (DGS-SLAM), a pioneering framework designed to effectively perform SLAM in dynamic environments by leveraging Gaussian Splatting. The paper identifies and addresses a significant limitation in existing dense SLAM systems, which typically assume static scenes, rendering them less effective in environments with dynamic objects. DGS-SLAM integrates Gaussian Splatting with a comprehensive filtering process, enhancing its robustness against photometric and geometric discrepancies introduced by moving objects.

Contributions and Methodology

DGS-SLAM introduces several key innovations:

1. Robust Mask Generation: The authors propose a novel method for mask generation that ensures photometric consistency across keyframes. This approach mitigates noise from segmentation inaccuracies and artifacts, thus improving dynamic object removal accuracy.
2. Loop-Aware Window Selection Mechanism: Employing unique keyframe IDs, this mechanism identifies loops between current and past frames, allowing for enhanced optimization of camera poses and the Gaussian map. This feature is crucial in maintaining map consistency over time.
3. Integration of Gaussian Splatting: The framework leverages Gaussian Splatting's strengths, such as differential rasterization, to enhance both scene reconstruction quality and rendering speed, demonstrating superior performance over traditional NeRF or SDF-based frameworks in dynamic settings.

The paper provides comprehensive details on the DGS-SLAM pipeline, divided into initialization, frontend tracking, and backend mapping processes. Gaussian splatting is utilized as the map representation, where each Gaussian is parameterized by its color, position, covariance, and opacity. Throughout the SLAM process, DGS-SLAM performs camera pose optimization, Gaussian map reconstruction, and dynamic element filtering.

Experimental Evaluation

The framework's performance is rigorously evaluated using standard SLAM benchmarks (TUM RGB-D dataset and Bonn RGB-D dataset). DGS-SLAM notably achieves state-of-the-art results in camera tracking and novel view synthesis in dynamic environments.

• Camera Tracking: DGS-SLAM exhibits more accurate pose estimation compared to other Gaussian Splatting and radiance field-based SLAM methods. It also outperforms traditional dynamic SLAM systems in several tested scenarios, with improvements in ATE and standard deviation metrics.
• Mapping and Rendering Quality: The paper reports superior rendering quality for novel view synthesis compared to contemporaneous approaches, confirmed by metrics such as PSNR, SSIM, and LPIPS. This affirms DGS-SLAM's competence in robustly handling dynamic elements while maintaining high-fidelity scene reconstruction.

Practical and Theoretical Implications

Practically, DGS-SLAM broadens the applicability of dense SLAM systems to dynamic, real-world environments, including robotics, autonomous navigation, and virtual or augmented reality systems. Theoretically, DGS-SLAM's introduction of Gaussian Splatting as a dynamic-aware SLAM representation opens avenues for further research, particularly in optimizing SLAM performance under variable environmental conditions.

Potential Future Directions

Potential future research directions include optimizing the computational efficiency of DGS-SLAM, given that online semantic segmentation can be resource-intensive. Exploring alternative dynamic filtering and segmentation techniques may further enhance robustness and efficiency. Moreover, extending the application of Gaussian Splatting SLAM to outdoor, large-scale environments could provide additional insights into its scalability and adaptability.

In summary, DGS-SLAM presents a significant step forward in the development of dynamic SLAM frameworks, demonstrating notable advancements in handling dynamic environments using Gaussian Splatting. This work sets a benchmark for future exploration into dynamic scene understanding and dense SLAM performance optimization.