改进的ORB特征提取及描述算法

Improved ORB Feature Extraction and Description Algorithm

针对传统ORB特征点提取分布密集的问题提出均匀化orb特征点提取以及针对经典二进制描述符鲁棒性差,特征点相似度高的问题提出使用局部差异二值(LBD)描述子,实现特征提取及匹配。首先构造图像金字塔,在金字塔每层使用四叉树的方法均匀提取特征点;之后采用LBD描述子,对特征点及其邻域采用网格的形式得到灰度和梯度信息,比较网格间的灰度和梯度信息;在提升辨别能力的同时,继承二进制描述子运行速度快和低存储的特点,最后在均匀化ORB特征点和LBD描述子的基础上对图像特征点进行提纯,匹配操作。实验结果表明,该方法计算的特征点更加均匀,区分度高,速度快,为后续的其他图像算法的应用提供了更加准确的匹配数据。

Aiming at the problems of dense distribution and uneven distribution of traditional ORB feature points extraction and the poor robustness and high similarity of feature points in classical binary descriptors, the homogenized ORB feature points extraction and its combination with local difference binary(LBD) descriptor were proposed to achieve feature extraction and matching. Firstly, an image pyramid is constructed and the quadtree method is used to extract feature points evenly in each layer of the pyramid. Then LBD descriptor was used to obtain the gray scale and gradient information of feature points and their neighborhoods in the form of grids, and the gray scale and gradient information between grids were compared. At the same time of improving the discrimination ability, the characteristics of fast running speed and low storage of binary descriptors are inherited. Finally, on the basis of homogenizing ORB feature points and LBD descriptor, the image feature points are purified and matched. Experimental results show that the feature points calculated by this method are more uniform, higher discrimination and faster, which provides more accurate matching data for subsequent applications of other image algorithms.