针对旋转不变性二进制描述算法(Oriented Fast and Rotated Brief, ORB)的尺度旋转性配准误差大, 配准率较低及随机采样一致性(Random Sample Consensus, RANSAC)算法随机性强且不稳定的问题, 提出一种ORB与RANSAC结合的快速特征匹配算法。首先, 对特征点提取方式进行优化选择, 消除特征边缘影响。之后构建简化的金字塔式尺度空间模型, 改进分层图像的尺度空间结构, 减少生成图像层数和数目; 然后采用梯度方向改进传统ORB算法中的主方向提取模式, 提高特征角点主方向的准确性。最后, 通过构建分块随机取样检测的方式改进RANSAC算法, 提高RANSAC算法的稳定性和图像配准的准确性。实验结果表明改进后的ORB和RANSAC融合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算法高, 尺度配准精度提高55.41%, 旋转配准精度提高26.66%。满足复杂图像快速精确配准拼接的精度和实时性要求。

In the binary description algorithm (Oriented Fast and Rotated Brief, ORB), scale and rotation cause a great error in the registration, and the registration rate is low. Meanwhile, the RANdom Sample Consensus (RANSAC) algorithm has an instability issue. Therefore, in this study, a fast feature matching algorithm was presented based on ORB with RANSAC. First, the feature point extraction method was optimized to eliminate the influence of feature edges. After constructing a simplified pyramid scale-space model, the scale-space structure of the layered image was improved by reducing the number of generated image layers. Subsequently, the gradient direction was used to improve the main direction extraction mode of the traditional ORB algorithm, and the accuracy of the main direction of the feature angular point was improved. Finally, the RANSAC algorithm was improved by applying block random sampling, which improved the stability and accuracy of image registration. Experimental results reveal that the improved ORB and RANSAC fusion algorithm performance greatly improved in terms of scale and rotation registration, and higher registration precision is exhibited in comparison with traditional ORB. The scale registration accuracy is improved by 55.41%, and the rotational registration accuracy is improved by 26.66%. These results indicate that the proposed algorithm basically meets the accuracy and real-time requirements for fast and accurate registration of complex images.