【作者】 宋迎新； 王文光； 杨予昊； 孙忠胜；

【Author】 SONG Yingxin;WANG Wenguang;YANG Yuhao;SUN Zhongsheng;School of Electronic Information Engineering, Beihang University;National Key Laboratory of Radar Detection and Sensing, Nanjing Research Institute of Electronics Technology;Millimeter-Wave Sensing and Intelligent Surveillance Research Platform, Hangzhou Innovation Institute of Beihang University;

【机构】 北京航空航天大学电子信息工程学院； 南京电子技术研究所雷达探测感知全国重点实验室； 北京航空航天大学杭州创新研究院毫米波感知与智能监控研究平台；

【摘要】 神经辐射场在新颖视角合成中表现优越,但对于前向视角为主的室内环境表现不佳。为了解决这些问题,提出了一种融合激光雷达与单目相机的神经辐射场方法,通过尺度对齐与位姿匹配实现精确融合,得到准确尺度并滤除错误位姿。点云作为几何先验引入辐射场,以构建体素与级联哈希表混合的占用结构跳过场景空白区域,并以点云深度监督训练,提升收敛速度并抑制漂浮伪影。实验表明该方法可以赋予辐射场准确的尺度信息,在新颖视图合成和深度估计方面取得了先进的性能。更多还原

【Abstract】 Neural radiance fields(NeRFs) excel in novel view synthesis but perform poorly in indoor scenes with predominantly forward-facing views. To address this problem, this paper proposes a NeRF method integrating LiDAR and monocular camera data. Scale alignment and pose registration enable accurate fusion, providing precise scale and filtering incorrect poses. LiDAR point clouds are incorporated as geometric priors using a hybrid voxel-hash occupancy structure to skip empty regions. Depth supervision from the point cloud accelerates convergence and suppresses floating clouds. The experimental results demonstrate that the proposed method provides accurate scale information to the radiance field, and achieves advanced performance in novel view synthesis and depth estimation.更多还原