Abstract: To solve the problem of large scale registration error and low rate of registration in binary description algorithm, this paper proposes an improved ORB feature matching algorithm with scale and direction. The algorithm uses the binary description algorithm ORB as the basis to construct a pyramid scale space, improve the scale space structure, simplify the number of scale space layers and the number of sampled images, make the process of extracting feature points more efficient, and Harris function is used to detect features, eliminate the influence of edge feature points, and extract feature points with scale information. Then, the method of gradient direction statistics is used to replace the way of computing the main direction by the gray-scale centroid method in the traditional ORB algorithm. The main direction of the neighborhood range is optimized and the accuracy of the image feature main direction is improved. The experimental results show that the improved ORB algorithm greatly improves the performance of scale and rotation registration, and the accuracy of registration is higher than the traditional ORB, and it can better meet the requirements of fast and accurate registration of complex images.

Key words: feature matching, Oriented fast and Rotated Brief（ORB） algorithm, scale space, feature main direction, gradient direction

摘要： 针对二进制描述算法（Oriented fast and Rotated Brief，ORB）尺度性配准误差大，配准率低的问题，提出一种尺度和方向改进的ORB特征匹配算法。该算法以二进制描述算法ORB为基础，构建金字塔式尺度空间，改进尺度空间结构，简化尺度空间层数和采样图像数目，使提取特征点的过程更加效率，并采用Harris函数检测特征，消除边缘特征点的影响，提取具有尺度信息的特征点；然后采用梯度方向统计方法改进传统ORB算法中通过灰度质心法计算主方向的方式，优化求解主方向邻域范围，以提高图像特征主方向的准确性。实验结果表明，改进后的ORB算法在尺度和旋转配准方面性能有很大提高，并且配准的精度较传统ORB更高，更能满足复杂图像快速精确配准的要求。

关键词: 特征匹配, 二进制特征描述算法（ORB）, 尺度空间, 特征主方向, 梯度方向