摘要
针对神经辐射场(NeRF)技术虽在三维重建复杂建筑场景中面临训练耗时长、细节描述有限等挑战,提出Structure-NeRF方法,使用无人机航拍获取多视图传感数据,结合多采样、减权及损失函数优化以减少锯齿,并引入几何约束(法线一致性、平面拟合和垂直/水平约束)提升重建精度。此外,利用曼哈顿世界模型(MWM)精确提取建筑主体结构与尺寸。实验表明,Structure-NeRF的峰值信噪比(PSNR)达36.23 dB,训练时间缩短了50%,结构相似性指数(SSIM)较主流模型mip-NeRF和iNGP分别提升13.41%和50.00%,PSNR分别提高10.45%和25.10%。Structure-NeRF显著优化了训练效率与渲染质量,实现了高精度的建筑空间模型三维重建。
To address the challenges faced by neural radiance field(NeRF)technology in the 3D reconstruction of complex architectural scenes,such as long training time and limited detail description,the Structure-NeRF method is proposed.Multi-view sensor data are acquired using unmanned aerial vehicle aerial photography.This method combines multi-sampling,weight reduction,and loss function optimization to reduce aliasing.Geometric constraints(normal consistency,plane fitting,and vertical/horizontal constraints)are introduced to improve reconstruction accuracy.Furthermore,the Manhattan World Model(MWM)is used to accurately extract the main structure and dimensions of buildings.Experimental results show that Structure-NeRF achieves a peak signal-to-noise ratio(PSNR)of 36.23 dB,reduces training time by 50%,and its structural similarity index(SSIM)is improved by 13.41%and 50.00%compared with the mainstream models mip-NeRF and iNGP,respectively,while its PSNR is increased by 10.45%and 25.10%,respectively.Structure-NeRF significantly improves training efficiency and rendering quality,achieving high-precision 3D reconstruction of architectural spatial models.
作者
吴建
常李霞
WU Jian;CHANG Lixia(School of Design,Shanghai Jiao Tong University,Shanghai 200090,China;Department of Information Technology,Shanxi Professional College of Finance,Taiyuan 030008,China)
出处
《传感器与微系统》
北大核心
2025年第11期76-81,共6页
Transducer and Microsystem Technologies
基金
国家自然科学基金面上项目(51278292)
上海交通大学文理交叉研究项目(17JCYA05)。
