改进了传统的尺度不变特征变换(SIFT)算法, 使其在进行图像匹配的同时, 可以求取出物体的旋转角度。首先, 利用SIFT特征对旋转保持不变的特性, 按照原算法提取出旋转前后两幅图像的SIFT特征, 分析特征点主方向的计算过程, 记录每个特征点主方向的角度值进行特征匹配。然后, 计算出每对匹配的SIFT特征点的主方向角度之差, 得到特征点的旋转角度; 采用迭代自组织聚类的方法分析得到的特征点旋转角度数据, 依据类内方差和类内样本数目, 选取正确的样本类。最后, 选用该样本类的均值作为物体的最终旋转角度。实验结果表明, 该方法在图像畸变不大时的误差在3°以内, 即使在部分遮挡的情况下, 也能较好地计算出旋转角度。在时间复杂度增加不大的情况下, 使SIFT算法具有了计算旋转角度的功能, 拓宽了应用方向。

The Scale-invariant Feature Transform(SIFT) algorithm was improved in this paper, which could match two pictures and could also compute the object rotation angles in the pictures. Firstly, the SIFT was used to extract two images according to the SIFT feature invariance. Then, the computing process for the main direction of the feature point was analyzed, and the main angle for every key point was recorded. After matching two pictures, the angle difference of main direction for each pair of matched SIFT feature points was calculated and the rotation angle of feature point was obtained. Afterthat, all of the rotation angles of the feature points were analyzed by iterative self-organizing clustering method. Finally, the correct class of samples was selected by the variance and the number of samples within the classe, and the mean of the correct class was used as the final rotation angle of the object. The experiment results indicate that the rotation angle error is within 3° when the image distortion is not significant and it can also well estimate the rotation angle even if the object is partially occluded. Furthermore, in the case of the time complexity does not increase obvionsly, the SIFT can compute the rotation angle of the object, which expands its applications.