基于特征点匹配的全景相机图像拼接方法研究
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

基金项目:

国家自然科学基金项目(31571570)、国家重点研发计划项目(2017YFD0700400-2017YFD0700403)和北京农业信息技术研究中心开放项目(KF2018W002)


Panoramic Camera Image Mosaic Method Based on Feature Points
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    全景相机可获取农机周围360°范围内的图像信息,具有覆盖范围大等特点,但需要对多镜头获取的图像进行拼接与融合,才能生成全景图像,为农机避障提供支持。以雷沃欧豹拖拉机为试验平台,搭载全景相机,获取实验农场的农田图像数据。首先对多幅图像进行预处理,包括通过柱面投影变换统一坐标系,采用基于特征点的SIFT算法提取图像的特征点并进行匹配;针对传统SIFT算法存在错误匹配而影响图像拼接质量的问题,使用RANSAC算法进行多次优化迭代,达到剔除错误匹配点的效果;针对匹配后生成的图像变换矩阵,为防止其线性结果不稳定并进一步优化结果,采用非线性的LM算法进行优化,使用线性加权平滑算法对图像进行融合,实现全景图像的生成。试验采用计算图像重叠区域相关系数定量评价图像拼接效果,并对获取的30组共60幅图像采用RANSAC算法和LM算法进行处理。结果表明,经过RANSAC算法处理后,误匹配点得到明显剔除,匹配特征点之间的平均几何距离偏移量明显减小,其平均值由39.4013像素下降至0.5819像素,相关系数由0.2878上升至0.7249。与手动设置阈值的剔除误匹配点方法进行了比较,经过RANSAC算法处理后的平均相关系数为0.7249,大于阈值设为0.4时的0.5933,以及阈值设为0.6时的0.2007,证明该算法能够实现多种情况下的图像拼接,剔除误匹配点;经过LM算法处理后,平均几何距离偏移量由0.5819像素进一步下降至0.5693像素,平均相关系数由0.7249进一步上升至0.7261,证明图像变换矩阵得到进一步优化,全景图像的拼接质量得到进一步提高。

    Abstract:

    The panoramic camera can obtain image information, within the scope of agricultural machinery around 360° coverage, with a large coverage and other characteristics, more conducive to agricultural machinery automatic navigation and obstacle avoidance. However, it is necessary to conduct image mosaic and fusion of the images acquired by the multilens as the test platform, so as to generate panoramic images for the research on obstacle avoidance of agricultural machinery. With a Lovol tractor for test platform, with a panoramic camera, images of the farmland in the experimental field were obtained, first of all, image preprocessing, mainly for cylindrical projection transformation unified coordinate system, after the SIFT algorithm based on feature points extraction of image feature points and matching for traditional SIFT algorithm matching errors and problems affecting the quality of image matching, using RANSAC algorithm, several optimization iteration to eliminate the effect of error matching points, in view of the image transformation matrix matching generated after, to prevent the instability of its linear results and further optimization results, the nonlinear LM algorithm was used to refine the image, and then the linear weighted smoothing algorithm was used to fuse the image to achieve the generation of panoramic image. Calculated by using the experimental image overlap coefficient of correlation effect of quantitative evaluation of image mosaicking, and 30 groups of image, a total of 60 images were processed by RANSAC algorithm and the LM algorithm, and the experimental results showed that after RANSAC algorithm processing, matching point obviously by mistake, the average geometric distance between the matching feature points offset was decreased significantly, fell by an average of 39.4013 pixels to 0.5819 pixels, the correlation coefficient was from 0.2878 to 0.7249, compared with the traditional method of removing mismatched points by manually setting the matching threshold of SIFT algorithm, the correlation coefficient of 0.7249 processed by RANSAC algorithm was obviously larger than that of 0.5933 when the threshold was set at 0.4 and 0.2007 when the threshold was set at 0.6, which proved that this study can be applied to image mosaic in many cases and eliminate mismatched points; after LM algorithm processing, the average geometric distance offset was further decreased from 0.5819 pixels to 0.5693 pixels, and the correlation coefficient was further increased from 0.7249 to 0.7261, proving that this study can further optimize the transformation matrix and improve the mosaic quality of panoramic images.

    参考文献
    相似文献
    引证文献
引用本文

徐弘祯,李世超,季宇寒,曹如月,张漫,李寒.基于特征点匹配的全景相机图像拼接方法研究[J].农业机械学报,2019,50(Supp):150-158.

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2019-04-25
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2019-07-10
  • 出版日期: 2019-07-10