Machine vision 机器视觉 学习笔记(4)Imaging Geometry成像几何

Transformation Between Object and Image Coordinates

校准是确定转换参数的过程

Least Squares Approximation最小二乘法近似

Suppose a system ax = b has N equations, N > 1, in only one unknown. (Equation 1.1)

One possibility of solving this inconsistent equation is to determine x from a part of the system, and ignore the rest.

A more reasonable method is to determine x that minimizes the average error in the N equations.

Sum of squares is a popular method to define average errors (Equation 1.2)

If there is an exact solution, minimum E=0.

If b is not proportional to a, the function E2 is a parabola with its minimum at the point defined by Equation 1.3.

Solving for x, Equation 1.4 is the least squares solution of the system

General formula: given any a not equal to 0 and any b, the error E is the length of the vector ax-b (Equation 1.5)

The parabola (Equation 1.6) has a minimum at the point defined by Equation 1.7.

The least squares solution to ax=b can be determined using Equation 1.8.

For an inconsistent system Ax=b with N equations and M unknowns. N is not equal to M.

The least squares solution is given by Equation 1.9.

If the matrix ATA is invertible, Equation 1.10 is the unique least squares solution.

Suppose we do a series of experiments, and expect the output y to be approximated by a linear function of t, y= C +Dt.

Example, measure the strain y of a linear elastic material subjected to a load t. If there is no error, only two measurement of ( y,t) are sufficient. But there will be experimental errors, a series of more than two sets of ( y,t) measured from the experiments are unlikely to fall on a straight line. We need to “average” all the experiments to determine the optimal line.

#线性近似，只测两个点就可以，测多了反而不在一条直线上了

 In practice, the assumption in the example of image guided medical robot that all target points are on a single horizontal work plane may not hold.

 

A typical 2D image from a single camera does not have sufficient information to recover the 3D information of the object.

Homogeneous Coordinates同质坐标

 A homogeneous vector corresponds to a straight line through the origin.

Advantages of Homogeneous Coordinates:

 Translation and perspective transformation can be expressed in matrix form

 Consecutive geometric transformation can be expressed as series of matrices with matrix multiplication

#平移和透视变换可以用矩阵形式表示/ 连续的几何变换可以用矩阵乘法表示为一系列的矩阵

To convert from n x 1 vector in homogeneous coordinates to coordinate of dimension n-1, we have to divide by the nth element and then delete the nth