【摘要】 在计算机视觉中,模型参数的鲁棒性估计是一个核心问题.虽然人们在这方面做了大量的研究,但是至今仍没有一个鲁棒性估计方法可以做到完全消除外点对模型估计的影响.当数据中外点比例很大的时候,传统随机抽样一致性(RANSAC)算法就需要大量的采样样本,这在实际运行时会降低算法效率.为了解决这个问题,提出了一个有效的基于采样的算法,利用射影变换下共线4点的交比不变性,对含有大量外点的数据进行筛选,剔除其中可能的外点数据,从而充分减少模型估计时的迭代次数.实验结果表明,当200个对应匹配中含有90%外点时,使用交比不变性约束进行平面单应矩阵估计,能够使最小迭代次数由69 074减少到11.由此可知,改进算法很大程度上提高了RANSAC的运算效率和对高噪声数据的处理能力.更多还原

【Abstract】 In computer vision,the robust estimation of model parameters is one of the essential tasks.Although a considerable amount of research has been done in this area,it is still hard to completely eliminate the impact of out-liers.When the fraction of outliers is significant,the traditional random sample consensus(RANSAC)algorithm re-quires a large number of samples and thus is prohibitively expensive.To tackle this problem,the invariance of cross ratio under perspective transformation was exploited for robust model estimation in the presence of large fraction of outliers.This approach will reject most of possible outliers existing among the data,thus significantly reducing the number of iterations needed for estimating the model.The experimental results show that,in the case of 180 outliers out of 200 matches,when cross ratio constraint is used for estimating planar homograph,the lower bound on the number of iterations decreases from 69 074 to 11.The improved algorithm is more efficient and enhances the ability of RANSAC in handling data with high noises.更多还原