| 摘 要: 3D高斯溅射(3DGS)的高效渲染能力显著推动了同时定位与建图(SLAM)的发展。然而现有方法在地图优化阶段通常依赖共可见性或随机采样策略，导致稀疏区域重建不足且对关键帧的选择针对性不足，同时缺少在真实生活环境下适用性验证。针对上述问题，提出了一种基于损失感知与迭代自适应的关键帧映射策略，通过动态计算候选帧的混合概率进行选择，从而提升地图重建质量与位姿估计的鲁棒性。在Replica和ScanNet数据集上的实验结果表明，相较于现有基于神经辐射场及3DGS的SLAM系统，跟踪精度与渲染质量均取得了明显提升。此外，通过自建数据集进行实验进一步验证了方法的可用性。 |
| 关键词: 3D高斯溅射 同时定位与建图 混合概率 真实生活环境 |
|
中图分类号:
文献标识码:
|
|
| 3D Gaussian Splatting SLAM Method Based on Loss-Aware Optimization |
|
Lv Xinyang1, Liu Xiaoping2
|
1.Zhejiang Sci-Tech University;2.Carleton University
|
| Abstract: The efficient rendering capability of 3D Gaussian Splatting (3DGS) has significantly advanced the development of simultaneous localization and mapping (SLAM). However, existing methods typically rely on visibility-based or random sampling strategies during map optimization, which results in insufficient reconstruction in sparse regions and insufficiently targeted selection of keyframes, while lacking validation in real-world environments. To address these problems,we present a loss-aware and iterative adaptive keyframe mapping strategy, which dynamically computes the hybrid probability for candidate frames for selection, thereby improving map reconstruction quality and the robustness of pose estimation. Experimental results on the Replica and ScanNet datasets demonstrate that, compared with state-of-the-art NeRF and 3DGS-based SLAM systems, the proposed method achieves notable improvements in tracking accuracy and rendering quality. Furthermore, experiments conducted on a self-collected real-world dataset confirm the practical effectiveness of the approach. |
| Keywords: 3D Gaussian Splatting SLAM Hybrid probability Real-world dataset |
