TL;DR
This paper introduces BDGS-SLAM, a novel framework that enhances 3D Gaussian Splatting (3DGS) for Simultaneous Localization and Mapping (SLAM) in dynamic environments. By integrating Bayesian filtering and multi-view probabilistic updates, BDGS-SLAM robustly handles moving objects, significantly reducing mapping errors and tracking drift compared to existing methods, leading to more accurate and artifact-free scene reconstructions.
What Problem Does This Paper Address?
Simultaneous Localization and Mapping (SLAM) is crucial for robots and augmented reality systems, enabling them to build maps of their surroundings while simultaneously tracking their own position. Recent advancements in 3D Gaussian Splatting (3DGS) have enabled highly realistic and fast scene reconstruction, leading to the development of 3DGS-based SLAM systems. However, these systems struggle in real-world environments where dynamic objects (like people or moving vehicles) are present. Traditional SLAM methods often assume static environments, and when faced with dynamic elements, they can suffer from "ghosting" artifacts, blurred textures, and inaccurate pose estimation. Existing 3DGS-SLAM methods often fail to effectively filter out or model these dynamic objects, leading to corrupted maps and unreliable tracking. This paper aims to bridge this gap by proposing a SLAM system that is specifically designed to be robust in dynamic environments.
Methods
BDGS-SLAM tackles the challenge of dynamic environments through a tightly coupled tracking and mapping framework that leverages three key probabilistic reasoning modules:
- Bayesian Filtering for Dynamic Verification: During the tracking phase, semantic information from YOLOv5 is used to detect dynamic objects. This semantic information, along with geometric features of Gaussian points, is fed into a Bayesian filter. This filter recursively estimates the probability of each Gaussian point being static or dynamic, providing a probabilistic margin rather than a hard classification. This helps to mitigate misclassifications and reduce the risk of incorrectly culling static points. An exponential decay factor is incorporated to ensure temporal consistency and smooth transitions in state estimation.
- Multi-View Probability Fusion: To further enhance the robustness of dynamic/static labeling, a multi-view probabilistic update mechanism is employed in the mapping phase. This module aggregates observations from co-visible keyframes. By using temporal weights (with an exponential decay factor), it prioritizes recent observations while still considering historical data. This approach helps to recover static Gaussians that might have been mistakenly identified as dynamic due to transient occlusions or noisy detections.
- Adaptive Dynamic Gaussian Optimization: Instead of hard-removing dynamic Gaussians, which can lead to map fragmentation, BDGS-SLAM uses a soft penalization strategy. This adaptive optimization approach applies constraints to suppress the influence of dynamic Gaussians during rendering and loss computation. This preserves the integrity of static scene structures and avoids introducing visual artifacts, while still allowing for temporal re-evaluation of ambiguous points.
The system also utilizes a Gaussian pyramid for hierarchical training, optimizing coarse geometry first before refining fine details, which improves efficiency and prevents overfitting.
Datasets
The proposed BDGS-SLAM system was evaluated on several publicly available and challenging datasets:
- TUM RGB-D Dataset: A standard benchmark for SLAM systems, providing ground-truth trajectories and sequences with both static and dynamic elements.
- BONN RGB-D Dynamic Dataset: This dataset features highly dynamic content with multiple fast-moving objects, making it ideal for testing robustness in aggressive dynamic scenarios.
- OpenLoris-Scene RGB-D Dataset: This dataset presents an exceptionally challenging evaluation benchmark with cluttered indoor settings, frequent lighting changes, diverse dynamic objects, and non-repetitive motion trajectories, assessing generalization and stability in real-world, long-term interactions.
Key Insights and Ideas
- Probabilistic Dynamic Labeling is Key: Moving beyond hard classifications, using Bayesian filtering to assign dynamic/static probabilities to Gaussian points is crucial for robustly handling uncertainty in dynamic scenes.
- Multi-View Evidence for Robustness: Fusing evidence from multiple viewpoints significantly improves the confidence in dynamic/static labels, especially in complex scenarios with occlusions.
- Soft Penalization Over Hard Removal: Instead of discarding potentially dynamic points entirely, adaptively penalizing their contribution during optimization and rendering is more effective for preserving map integrity and avoiding artifacts.
- Tightly Coupled Framework: Integrating semantic detection, probabilistic filtering, and adaptive optimization within a single, tightly coupled framework is essential for efficient and accurate performance in dynamic environments.
Key Results
BDGS-SLAM demonstrated state-of-the-art performance across the evaluated datasets:
- Localization Accuracy: On the TUM RGB-D dataset, BDGS-SLAM achieved the lowest average ATE (0.0247 m), a significant improvement over baseline methods. It also showed superior performance on the challenging BONN RGB-D dataset, maintaining stable tracking even in sequences where other methods failed.
- Reconstruction Quality: BDGS-SLAM consistently produced higher-fidelity reconstructions, evidenced by improved PSNR (e.g., 23.54 dB on TUM) and SSIM (e.g., 0.900 on TUM), and lower LPIPS scores. Visual inspection confirmed fewer ghosting artifacts and cleaner static scene reconstructions.
- Runtime Efficiency: The system achieved competitive runtime performance, with its tracking module being the fastest among evaluated methods. The overall average runtime was approximately 565.3 ms per frame, enabling near real-time operation.
- Ablation Studies: Experiments confirmed the significant contribution of each proposed module (Bayesian filtering, multi-view fusion, and adaptive optimization) to the overall performance.
Limitations and Future Work
- Limitations: While BDGS-SLAM significantly improves robustness in dynamic environments, the paper notes that the computational cost of semantic segmentation (YOLOv5) is excluded from reported timings, as it runs asynchronously. Further acceleration could be achieved by using lighter object detectors or pruning low-confidence dynamic labels earlier.
- Future Work: The authors plan to enhance the system's real-time capability and robustness by integrating lightweight semantic segmentation networks and developing memory-efficient modules for long-term dynamic scene understanding. They also suggest future directions in probabilistic modeling, attention-based Gaussian filtering, and physically grounded motion priors.
Summary updated