nav emailalert searchbtn searchbox tablepage yinyongbenwen piczone journalimg journalInfo journalinfonormal searchdiv searchzone qikanlogo popupnotification paper paperNew
2026, 05, 17-27+106
面向交通基础设施的大场景三维重建算法研究
基金项目(Foundation): 国家自然科学基金重点项目(42430711)
邮箱(Email):
DOI: 10.3969/j.issn.2097-017X.2026.05.003
投稿时间: 2026-07-21
投稿日期(年): 2026
发布时间: 2026-09-15
移动端阅读
摘要:

针对交通基础设施大场景三维建模中覆盖范围广、场景密度分布不均、有限显存条件下训练 效率低,以及分块边界易产生几何与纹理接缝等问题,提出一种基于三维高斯溅射(3D Gaussian Splatting,3DGS)的分块融合重建算法。利用稀疏点云开展全局粗训练,构建道路走向、桥体轮廓 及周边地物的几何先验;依据高斯密度与八叉树空间关系自适应划分子块,结合视角‑分块关联度 筛选机制减少冗余训练影像,实现局部计算负载均衡;通过重叠区域几何与颜色一致性约束完成子 块融合,并采用多层次细节渲染提高大场景交互显示效率。实验采用无人机采集的 1892 张高分辨 率影像,覆盖约 3 km²的交通基础设施走廊复合场景,以道路和桥梁为主要建模对象,同时包含河 道、农田、房屋与楼宇等沿线环境。在相同条件下,相较运动恢复结构‑多视图立体(SfM‑MVS)基 线,本文方法将总处理时间从 267 min 缩短至 197 min,降幅达 26.2%;六类测试场景的平均结构相似性(SSIM)由 0.459 提升至 0.840,平均峰值信噪比(PSNR)由 17.041 dB 提升至 28.495 dB,平均学 习感知图像块相似度(LPIPS)由 0.589 降低至 0.291。结果表明,该方法兼顾平方公里级交通基础 设施场景的重建效率、结构连续性与纹理保真度,可为数字化建模、巡检成果归档及交通数字孪生可视化提供三维模型基础。

Abstract:

Large-scale 3D reconstruction of transportation infrastructure faces multiple challenges, including wide scene coverage, uneven scene point density, limited GPU memory, and seam artifacts at block boundaries. This paper proposes a block-fusion method based on 3D Gaussian Splatting (3DGS). The sparse point clouds are used for global coarse training to establish geometric priors for road alignments, bridge profiles, and surrounding scene features. The entire scene is adaptively partitioned based on Gaussian density and octree-based spatial relationships. Meanwhile, a view-block relevance filtering mechanism is introduced to eliminate redundant images and balance local computational loads. Geometric and color consistency constraints applied to overlapping regions enable seamless block fusion, and level-of-detail rendering improves the efficiency of interactive visualization. Experiments are conducted on 1892 UAV images capturing an approximately 3 km² transportation corridor, with roads and bridges as the primary reconstruction targets. Compared with the SfM-MVS base‑line method, the proposed method reduces total processing time from 267 min to 197 min, with a 26.2% reduction. Across six test categories, the average SSIM increases from 0.459 to 0.840, the average PSNR increases from 17.041 dB to 28.495 dB, and the average LPIPS decreases from 0.589 to 0.291. Experimental results verify that the proposed method improves reconstruction efficiency, structural continuity, and texture fidelity, and can support digital modeling and digital twin visualization of transportation infrastructure.

参考文献

[ 1] 交通运输部 . 交通运输部关于推动交通运输领域新型基础设施建设的指导意见(交规划〔2020〕75 号)[EB/OL]. (2020-08-06)[2026-07-14]. https://www. gov.cn/zhengce/zhengceku/2020-08/06/content_5532842.htm.

[ 2] 交通运输部 . 数字交通“十四五”发展规划(交规划发〔2021〕102 号)[EB/OL].(2021-12-22)[2026-09-12].https://xxgk.mot.gov.cn/jigou/zhghs/202112/t20211222_3632469.html.

[ 3] 刘博 . 倾斜摄影三维重建算法研究及可视化应用[D].桂林:桂林电子科技大学,2023.

[ 4] MILDENHALL B, SRINIVASAN P P, TANCIKM, et al. NeRF:representing scenes as neural radiance fields for view synthesis[C/OL]//Computer VisionECCV 2020. Cham: Springer,2020:405-421[2026-09-12]. https://arxiv.org/abs/2003.08934.

[ 5] BARRON J T, MILDENHALL B, VERBIN D, et al. Mip-NeRF 360: unbounded anti-aliased neural radiance fields[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). IEEE,2022:5460-5469.

[ 6] KERBL B, KOPANAS G, LEIMKUEHLER T, et al. 3D Gaussian splatting for real-time radiance field rendering[J]. ACM Transactions on Graphics, 2023,42(4):1-14.

[ 7] FENG G F, CHEN S Y, FU R, et al. FlashGS: efficient 3D Gaussian splatting for large-scale and highresolution rendering[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE,2025:26652-26662.

[ 8] 闫利,费亮,陈长海,等 . 利用网络图进行高分辨率航空多视影像密集匹配[J]. 测绘学报,2016,45(10):1171-1181.

[ 9] 王金虎,许俊辉,史嘉奇,等 . 基于无人机倾斜摄影的三维全景模型构建对比分析[J]. 建模与仿真,2022,11(6):1434-1444.

[10] LOWE D G. Distinctive image features from scaleinvariant keypoints[J]. International Journal of Computer Vision,2004,60(2):91-110.

[11] BAY H, TUYTELAARS T, VAN GOOL L.SURF: speeded up robust features[M]//Computer Vision‑ECCV 2006. Berlin, Heidelberg: Springer,

2006:404-417.

[12] SCHÖNBERGER J L, FRAHM J M. Structure-frommotion revisite[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR).IEEE,2016:4104-4113.

[13] KAZHDAN M, BOLITHO M, HOPPE H. Poisson surface reconstruction[C/OL]//Proceedings of the Fourth Eurographics Symposium on Geometry Processing. The Eurographics Association, 2006: 61-70.[2026-09-12]. https://doi. org/10.2312/SGP/SGP06/061-070.

[14] DU Y, SUN F C, LV X, et al. Advances in neural radiance fields for large-scale 3D scene reconstruction: a comprehensive review[M]//Theoretical Computer Science. Singapore: Springer Nature Singapore, 2025:123-139.

[15] LIN J Q, LI Z H, TANG X, et al. VastGaussian: vast 3D Gaussians for large scene reconstruction[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE,2024:5166-5175.

[16] 曹振中,光金正,张千一,等 . 基于 3D 高斯溅射的 3 维重建技术综述[J]. 机器人,2024,46(5):611-622.

[17] WU Y C, QI Z P, SHI Z W, et al. BlockGaussian: efficient large-scale scene novel view synthesis via adaptive block-based Gaussian splatting[EB/OL]. (2025-04-12). https://arXiv.org/abs/2504.09048.

[18] XU Q S, SU W J, LIAO X Y, et al. Geo-3DGS:multi-view geometry consistency for 3D Gaussian splat‑ting and surface reconstruction[EB/OL]. (2024-09-27)[2026-09-12]. https://openreview.net/forum?id=c4Nh4A8Xn5.

[19] LIU S Y, TANG X, LI Z H, et al. OccluGaussian:occlusion-aware Gaussian splatting for large scene reconstruction and rendering[C]//Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). IEEE,2025:26643-26652.

[20] 高建,陈林卓,沈秋,等 . 基于三维高斯溅射技术的可微分渲染研究进展[J]. 激光与光电子学进展,2024,61(16):1611010.

[21] YANG L H, KANG B Y, HUANG Z L, et al. Depth anything V2[C]//Proceedings of Advances in Neural Information Processing Systems 37. Neural Information Processing Systems Foundation, Inc. (NeurIPS) ,2024:21875-21911.

[22] CHUNG J, OH J, LEE K M. Depth-regularized optimization for 3D Gaussian splatting in few-shot images[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops

(CVPRW). IEEE,2024:811-820.

[23] TRIGGS B, MCLAUCHLAN P F, HARTLEY R I,et al. Bundle adjustment: a modern synthesis[C]//Vision Algorithms: Theory and Practice. Berlin, Heidelberg: Springer,2000:298-372.

[24] WANG Z, BOVIK A C, SHEIKH H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Transactions on Image Processing,2004,13(4):600-612.

[25] HORÉ A, ZIOU D. Image quality metrics: PSNR vs.SSIM[C]//Proceedings of 20th International Conference on Pattern Recognition. IEEE,2010:2366-2369.

[26] ZHANG R, ISOLA P, EFROS A A, et al. The unreasonable effectiveness of deep features as a perceptual metric[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE,2018:586-595.

基本信息:

DOI:10.3969/j.issn.2097-017X.2026.05.003

引用信息:

[1]胡彦卿,徐浩,张晨熹,等.面向交通基础设施的大场景三维重建算法研究[J],2026,(05):17-27+106.DOI:10.3969/j.issn.2097-017X.2026.05.003.

基金信息:

国家自然科学基金重点项目(42430711)

投稿时间:

2026-07-21

投稿日期(年):

2026

发布时间:

2026-09-15

检 索 高级检索

引用

GB/T 7714-2015 格式引文
MLA格式引文
APA格式引文