机器人2025，Vol.47Issue(5)：708-717,10.DOI:10.13973/j.cnki.robot.240130
RefN-SLAM:反射场景下的神经SLAM方法
RefN-SLAM:Neural SLAM Method in Reflective Scenes
摘要
Abstract
Traditional SLAM(simultaneous localization and mapping)methods rely on the assumption of photometric consistency,and often fail to handle scenes with complex lighting variations.Therefore,a reflection scene mapping method based on neural radiance field,termed RefN-SLAM,is proposed.Specifically,two neural radiance fields are used to model the high-light and low-light parts of the scene separately.The final scene color representation is obtained by weighting and summing them with fixed tone mapping coefficients.Meanwhile,the depth perception of the scene is further enhanced by combining surface-aware sampling and perspective-aware sampling,and the reconstruction accuracy and computational efficiency are improved through a coarse-to-fine optimization process.Finally,joint optimization of scene representation and camera pose is performed on a global keyframe pixel database.The experimental results demonstrate that RefN-SLAM method achieves satisfactory reconstruction performance in a chemistry laboratory setting and exhibits excellent tracking performance in both synthetic datasets and real-world robotic experiments.关键词
视觉SLAM(同步定位与地图构建)/神经辐射场/室内反射场景/复杂光照变化Key words
visual SLAM(simultaneous localization and mapping)/neural radiance field(NeRF)/indoor reflection scene/complex illumination change引用本文复制引用
胡佳豪,尚伟伟,汤新胜,张飞..RefN-SLAM:反射场景下的神经SLAM方法[J].机器人,2025,47(5):708-717,10.基金项目
中国科学院战略先导科技专项(XDB0450302) （XDB0450302）
国家自然科学基金(U22A2056). （U22A2056）
