基础矩阵的鲁棒性估计是计算机视觉领域的一个基本问题。为了提高基础矩阵的估计精度，首先指出了现有的鲁棒性算法——RANSAC和MLESAC理论上的缺陷和实际应用中的问题；然后通过详细分析局外点复杂的成因，同时运用混合高斯分布代替均匀分布分别对不同成因的局外点进行了有针对性的建模，并提出了一种鲁棒性更强的算法——GMSAC。实验结果表明，相比于MLESAC算法，GMSAC算法提供了更高的模型似然度和计算精度。

A new method is presented for robustly estimating the fundamental matrix from image correspondence. Starting from RANSAC and MLESAC algorithm, we addressed some problems posed from both practical and theoretical standpoints and propose a new algorithm GMSAC. GMSAC adopts the same sampling strategy and maximization likelihood theory as the previous approaches. But instead of uniform distribution used by MLESAC, GMSAC choose Gaussian mixture to model the outliers. Due to the complex nature of outliers’ occurrence, Gaussian mixture is more suitable for the distribution of outliers. We make a detailed analysis on the formation of outliers, and model different types of outliers respectively. Results are given for several image sequences, and it is demonstrated that this method gives results superior to the original MLESAC.