摘要:
首先研究探讨了基于绝对二次曲线(the absolute conic)进行摄像机自标定鲁棒性差的
内在原因.研究发现,该类方法鲁棒性不足的原因主要有三个方面:1)在目标函数的全局最小点
处存在大范围的平坦区域,使得任何数值优化算法难以达到全局最小点;2)当存在噪声时,上
述平坦区域内会出现大量局部极小值,这样数值优化算法就非常容易收敛到靠近初值的局部极
小值,使得算法对初始值的选取十分敏感;3)当有噪声时,目标函数的全局最小值极易偏离正
确值.这样,即使数值算法找到了全局最小值,该最小值也不再对应正确的摄像机内参数值.鉴
于上述情况,探讨了如何通过平面场景来确定内参数矩阵的初始值,而后进一步利用Kruppa方
程的约束来精化内参数矩阵的二步式方法.
Abstract:
It is well recognized that the IAC (Image of the Absolute Conic) based
camera calibration techniques are not quite robust, however, there are few reports on
its underlying reasons in the literature. In this paper, we find that the following three
sources largely contribute to the non-robustness of the IAC based techniques. 1)The
global minimum of the cost function lies on a large flat area, which makes any numerical
optimization methods difficult to reach it. 2)With noise, many local minimums
could appear in the above flat area, which makes any numerical optimization methods
quite sensitive to the choice of the initial point since the used method could converge
easily to the local minimum close to the initial point. 3) With noise, the point corresponding
to the global minimum of the cost function could deviate significantly from
the one sought for, thus it is difficult to get the real calibration parameters by minimizing
the cost function. In addition, we explore a two-step calibration technique. In this
new technique, firstly a plane, which is distant from the origin, is used to obtain an
initial solution, then Kruppa equations are used to refine this initial estimation. The
experiments on simulated data as well as on real images validate our new technique.