• 中国科学引文数据库(CSCD)来源期刊
  • 中文核心期刊中文科技核心期刊
  • Scopus RCCSE中国核心学术期刊
  • 美国EBSCO数据库 俄罗斯《文摘杂志》
  • 《日本科学技术振兴机构数据库(中国)》
二维码

隧道建设(中英文) ›› 2023, Vol. 43 ›› Issue (7): 1118-1126.DOI: 10.3973/j.issn.2096-4498.2023.07.004

• 研究与探索 • 上一篇    下一篇

联合SIFT-GPU和整体平差法的隧道影像拼接研究

张栋樑1, 宫志群1, 杨世廷1, 陆业宁1, 占静2, *, 安晓亚3, 4   

  1. 1. 中国建设基础设施有限公司, 北京 100029 2. 武汉天际航信息科技股份有限公司, 湖北 武汉 430074 3. 西安测绘研究所, 陕西 西安 710054; 4. 地理信息工程国家重点实验室, 陕西 西安 710054
  • 出版日期:2023-07-20 发布日期:2023-08-06
  • 作者简介:张栋樑(1979—),男,黑龙江牡丹江人,2008年毕业于同济大学,道路与铁道工程专业,博士,高级工程师,现从事道路工程技术管理工作。Email: 466716578@qq.com。*通信作者: 占静, Email: zhanjing@whulabs.com。

Tunnel Image Stitching Based on ScaleInvariant Feature TransformGraphics Processing Unit and Integral Adjustment Method

ZHANG Dongliang1, GONG Zhiqun1, YANG Shiting1, LU Yening1, ZHAN Jing2, *, AN Xiaoya3, 4   

  1. (1. China Construction Infrastructure Co., Ltd., Beijing 100029, China; 2. Wuhan Tianjihang Information Technology Co., Ltd., Wuhan 430074, Hubei, China; 3. Xian Research Institute of Surveying and Mapping, Xian 710054, Shaanxi, China; 4. State Key Laboratory of Geoinformation Engineering, Xian 710054, Shaanxi, China)
  • Online:2023-07-20 Published:2023-08-06

摘要:

针对当前隧道影像拼接过程中效率低、累积误差大的问题,研究提出一种联合SIFT-GPUscale-invariant feature transform-graphics processing unit)协同处理算法和整体平差法的隧道影像拼接技术。首先,基于车载移动测量技术获取隧道内部的影像和点云数据,利用点云数据和共线方程对原始影像进行纠正处理,并先后使用MASK匀光算法和Wallis滤波匀光法对影像进行匀光匀色; 然后,融合SIFT匹配算法和GPU-CPUcentral processing unit)协同处理技术加速提取所有待拼接影像特征点,使用RANSACrandom sample consensus)算法剔除误配点;最后,使用平移旋转缩放变换模型对400张影像进行拼接,采用整体平差算法对拼接结果进行误差补偿,利用GDALgeospatial data abstraction library)分块处理技术输出拼接成果。研究结果表明: 1)相较于原始的SIFT匹配算法,研究使用的SIFT-GPU协同处理算法使影像匹配效率提高了3~25倍; 2)整体平差算法较好地解决了随着待拼接影像的增加而引起的误差累积问题,相较于传统算法,误差降低至原来的1/3。

关键词: 隧道, 影像拼接, SIFT, RANSAC, 整体平差法

Abstract: The existing tunnel image stitching is inefficient and prone to large cumulative errors. Therefore, a technology for tunnel image stitching based on a collaborative processing algorithm of scaleinvariant feature transform(SIFT)graphics processing unit(GPU) and integral adjustment method is required. First, the image and point cloud data inside the tunnel are obtained using vehiclemounted mobile measurement technology; the original image is corrected using point cloud data and collinear equation, and the light and color are homogenized using the MASK smoothing algorithm and Wallis filter smoothing method. Then, the SIFT matching algorithm is integrated using a collaborative processing technology of the GPUcentral processing unit to extract feature points of all images to be stitched, and the random sample consensus algorithm is used to remove mismatches. Finally, a translation rotation scaling model is used to stitch 400 images, the global adjustment algorithm is used to compensate for the error, and the block processing technology of the geospatial data abstraction library is used to output the results. According to the experimental findings, the collaborative processing algorithm of SIFTGPU can improve imagematching efficiency by 3 to 25 times compared to the original SIFT matching algorithm. Compared to the original algorithm, the integral adjustment algorithm can more effectively address the error accumulation problem caused by the increase of the image to be stitched, resulting in a threefold reduction in error.

Key words: tunnel, image stitching, scaleinvariant feature transform, random sample consensus, integral adjustment