Exposure compensation and computational resource optimization in 3DGS-SLAM

Existing 3DGS-SLAM systems frequently exhibit compromised tracking performance due to disturbances from illumination variations, texture discontinuities, and other noise sources, while also facing progressively accumulating GPU memory demands in long-term sequential tasks. To mitigate these issues, this paper presents a robust SLAM framework based on 3D Gaussian Splatting (3DGS), which incorporates an adaptive exposure compensation mechanism to optimize per-frame exposure parameters during tracking, thereby stabilizing pose estimation under challenging photometric conditions. Furthermore, a hierarchical computational resource management strategy is proposed to dynamically adjust rendering parameters according to real-time GPU memory availability, effectively reducing computational overhead without sacrificing reconstruction quality. Extensive experiments on the TUM RGB-D and Replica datasets demonstrate that our method achieves significantly higher pose estimation accuracy compared to state-of-the-art approaches. Quantitative evaluations using PSNR, SSIM, and LPIPS metrics confirm superior reconstruction fidelity. Notably, the optimized implementation reduces GPU memory consumption by approximately 15%, validating its efficiency in resource utilization. These results collectively verify that the proposed algorithm maintains high localization precision and robustness in complex static environments while optimizing computational resource allocation.