结合特征定位噪声表征的单应矩阵精确鲁棒估计

针对基于特征匹配的单应矩阵估计方法的特征定位噪声的各向异性非同分布对其精度和鲁棒性的影响,提出了一种结合特征定位噪声表征的单应矩阵估计方法。该方法采用协方差矩阵来表征特征点定位噪声;基于协方差矩阵加权采样一致性(CWSAC)的内点检验方法来提高单应矩阵估计的鲁棒性。最后,提出一种单应矩阵高精度估计算法——协方差加权Levenberg-Marquardt(CW L-M)法。该方法结合协方差矩阵重新定义优化目标函数,提高了单应矩阵的估计精度。基于仿真数据和真实图像的实验表明,在相同定位噪声和内点比例条件下,本文算法的估计精度显著优于RANSAC(RANdom SAmple Consensus)、LMedS(Least Median of Squares),PROSAC(PROgressive SAmple Consensus)、M-SAC(M-estimator SAmple Consensus)和MLESAC(Maximum Likelihood SAmple Consensus)等传统算法,投影均方误差比次优方法降低了3%~21%。另外,本文方法对定位噪声和内点比例变化均具有较好的鲁棒性。

Accuracy and robust estimation of homography based on feature point location noise

The feature location noise from feature-based homography estimation methods is isotropic and non-identical distribution, and it effects the accuracy and robustness of homography estimation methods significantly in practical applications. Therefore, this paper proposes a high accuracy and robust homography estimation method based on location noise of feature points. The method uses a covariance matrix to characterize the location noise of feature points and takes an inner point verification method based on Covariance matrix Weight SAmple Consensus(CWSAC) to improve the robustness of the homography estimation method. Finally, a high accuracy homography matrix refined method, Covariance matrix Weight Levenberg-Marquardt(CW L-M) is proposed by combining covariance matrix with Levenberg-Marquardt method, and it improves the estimation accuracy of homography matrix by redefining a optimized object function. The experiments on simulation data and real images show that as compared with state-of-the-art methods, such as RANSAC(RANdom SAmple Consensus),LMedS(Least Median of Squares), PROSAC(PROgressive SAmple Consensus), M-SAC(M-estimator SAmple Consensus)and MLESAC(Maximum Likelihood SAmple Consensus), the accuracy of homography estimation has improved greatly and the root mean squares of reproject error has reduced 3%-21% than that of the subprime method in the same location noise and the same inlier proportion. In addition, the proposed method is robust to the noise level and inlier proportion changing.