基于特征点不确定性加权误差的位姿估计新方法

计算机视觉中一般是利用优化技术最小化目标函数来估计位姿,目标函数通常是由图像平面上特征点重投影误差构成,并且假设测量噪声是各向同性且独立同分布的高斯噪声,所得到的位姿是在该假设条件下的极大似然最优估计。然而,在实际应用中这种假设并不总是成立,测量噪声通常是各向异性且非独立同分布,而且常常具有很强的方向性。为此,本文提出了一种新的特征点位姿估计方法,首先对特征点的方向不确定性建模,然后将方向不确定性融入到重投影误差中,构造基于不确定性加权误差的新目标函数,最后利用Levenberg-Marquardt算法优化目标函数求解位姿。大量实验结果表明,本方法可以适应不同程度的方向不确定性,精度优于现有迭代方法。而且随着不确定性的增加,位姿解的精度并没有明显变差。

A new pose estimation method based on uncertainty-weighted errors of the feature points

Pose estimation problem in computer vision is often solved by minimizing cost functions that combine the reprojection errors in the 2D images,but the cost functions are based on the assumption that measurement noises are isotropic,independent and identically distributed(i.i.d.) at every 2D feature point,so the solution of pose becomes statistically meaningful and optimal in the sense of maximum likelihood.However,this assumption is rarely the case in real applications,where the positional noise not only varies at different features,but also has strong directionality.In this situation,we propose a new pose estimation method that incorporates the directional uncertainty of the points into reprojection errors and constructs a new uncertainty-weighted cost function.The method can adapt to varying degrees of directional uncertainty and guarantee more accurate results.In particular,unlike other reprojection-error-based methods,our method does not degrade with increase in directionality of uncertainty.Finally we demonstrate the effectiveness of the method by experiments on synthetic data that contains high directional uncertainty.