大豆冠层多源图像特征点配准方法研究

Registering feature points for multi-source images of soybean canopies

关键词:  大豆冠层  多源图像  PMD摄像机  彩色摄像机  配准

Key words:  soybean canopies  multi-source images  PMD camera  color camera  registration

摘要:

针对具有颜色信息的大豆冠层三维结构形态的重建问题，采用PMD摄像机与彩色摄像机相结合的多源图像采集系统获取大豆冠层多源图像，对大豆冠层多源图像特征点配准方法进行研究。以彩色图像和强度图像为研究对象，利用仿射变换实现彩色图像坐标系到PMD图像坐标系的转换；利用Harris算法检测图像特征点，采用基于归一化互相关系数法（NCC）实现特征点粗匹配。为克服传统RANSAC算法抽样次数较多及和数据检验时间较长的弊端，提出在特征点匹配阶段，按照可信度将特征点对排序，从可信度高的点对开始抽取的方法来优化经典RANSAC算法，进而实现特征点精匹配，最终完成多源图像特征点配准。为验证本研究提出的图像配准算法的有效性，将该算法与传统图像配准算法相对比，结果表明：室外和室内环境下，样本组的平准正确配准率分别为83%和87%，均优于传统图像配准算法，并满足快速配准大豆冠层多源图像特征点的要求。

Abstract:

In order to reconstruct the three-dimensional structure of soybean canopy with color information, a multi-source image acquisition system combining PMD camera and color camera was used to acquire the multi-source image of soybean canopy, and the feature points of multi-source images of soybean canopy were studied. By using color image and intensity images as research objects, affine transformation was exploited to realize the transformation from color image coordinate system to PMD image coordinate system firstly, Harris algorithm was then to test the image feature points. Normalized cross correlation was adopted to achieve coarse matching for feature points. In addition, for the purpose of overcoming disadvantages of classic RANSAC algorithm, such as more sampling times and longer time of data validation, an improved RANSAC algorithm was used to realize precise matching, which was based on classification by credibility during stage of matching feature points. We finally accomplished Feature point registration of multi-source image after finishing the steps mentioned above was accomplished. To verify the validity of the image registration algorithm proposed in this study, algorithms of this article were compared to with the traditional image registration algorithm. The test results showed that the correct registration ratios of the sample groups were 83% and 87%, respectively under the circumstance of outside and inside environments, and both of them were better than the classical image registration algorithms, which could meet the need of quick registering feature points of multi-source images of soybean canopies.