Abstract: In order to improve robustness of the traditional camera self-calibration algorithm based on Kruppa equations, the new method of camera self-calibration based on tensor voting is first proposed. The SIFT algorithm, based on scale invariant property, is adopted to extract and match the feature points of each image, which are taken from three different angles to the same scene. The first eight feature points with robustness are figured out with tensor voting and sorting. Then the fundamental matrix and the pole points are calculated by 8 points’ algorithm, and finally the parameter matrices of camera can be obtained by the Kruppa equations and the Structure From Motion（SFM） algorithm. Compared with other algorithms through the comparative experiments, this method is proven to be more accurate, meanwhile, it can be regarded as a new robust camera self-calibration algorithm by the simulation experiments with Gaussian noise.

Key words: camera self-calibration, Kruppa equations, Scale Invariant Feature Transform（SIFT）, tensor voting, fundamental matrix

摘要： 针对传统的基于Kruppa方程摄像机自标定算法的欠鲁棒性，首次提出将鲁棒的张量投票算法用于摄像机自标定方法中。利用基于尺度不变的SIFT算法查找并匹配出每对图像的特征点，其中待匹配图像由摄像机对同一场景从三个不同角度位置拍摄，对图像张量投票后按棒张量特征值降序排序，由此筛选得到具有鲁棒性边缘特征的前八对特征点，利用八点算法求解相应的基础矩阵和极点，根据Kruppa方程和三维重建（SFM）算法求得摄像机参数矩阵。实验结果证明，该方法具有较高标定精度，并通过加入高斯噪声的仿真实验证明该算法是一种鲁棒的摄像机自标定方法。

关键词: 摄像机自标定, Kruppa方程, 尺度不变特征变换（SIFT）, 张量投票, 基础矩阵