摘要
针对大规模场景重建中可扩展性与渲染质量的挑战,提出一种基于三维高斯溅射(3DGS)的分块重建框架。方法上,使用无人机(UAV)航拍传感器采集场景图像数据,通过场景分块、可见性相机选择与渐进式点云扩展提升可扩展性;在渲染方面引入光线—高斯交叉与高斯密度控制提升效率,采用基于ConvKAN的外观解耦模块缓解光照不均,并结合颜色、深度失真与法向一致性损失优化训练。结果在Mill—19、UrbanScene3D和MatrixCity数据集上显示,本方法在渲染质量与精度上均优于现有主流方法,并通过自采集无人机视频验证了强泛化能力。结论表明,该框架能够高效重建大规模场景并实现高保真新视图合成。
To address the challenges of scalability and rendering quality in large-scale scene reconstruction,a block-based reconstruction framework based on three-dimensional Gaussian splatting(3DGS)is proposed.Methodologically,unmanned aerial vehicle(UAV)aerial photography sensors are used to capture scene image data.Scalability is improved through scene blocking,visibility camera selection,and progressive point cloud expansion.In terms of rendering,ray-Gaussian intersection and Gaussian density control are introduced to improve efficiency.A ConvKAN-based appearance decoupling module is used to alleviate uneven illumination.Training is optimized by combining color distortion and depth distortion with normal consistency loss.Results on datasets such as Mill—19,UrbanScene3D,and MatrixCity show that this method outperforms existing mainstream methods in both rendering quality and accuracy.Its strong generalization capabilities are also demonstrated using self-collected unmanned aerial vehicle videos.Conclusions demonstrate that this framework is capable of efficiently reconstructing large-scale scenes and achieving high-fidelity new view synthesis.
作者
付瀚思
陈娜
FU Hansi;CHEN Na(School of Art,Henan Forestry Vocational College,Luoyang 471002,China;School of Civil Engineering and Architecture,Henan University of Science and Technology,Luoyang 471023,China)
出处
《传感器与微系统》
北大核心
2025年第12期125-129,共5页
Transducer and Microsystem Technologies
基金
2025年度河南省高校人文社会科学研究一般项目(2025-ZDJH-162)
2026年度河南省高校人文社会科学研究一般项目(2026-ZDJH-475)。
关键词
无人机航拍传感器
大规模场景
三维重建
UAV aerial photography sensor
large-scale scene
three-dimensional reconstruction
