在三维空间中,如何通过一组离散的3维坐标来拟合平面空间,使得任意一组空间坐标**(Px, Py, Pz)**在平面方程上。
给定n个三维点的坐标,根据这些点坐标由最小二乘法拟合平面。
平面方程一般为Ax+By+Cz+D = 0,为了方便计算这里将平面方程变形为z=Ax+By+C;
像素坐标通过相机内参及外参数矩阵可以得到n个三维点坐标,理想情况下应满足
{ z 0 = A x 0 + B y 0 + C z 1 = A x 1 + B y 1 + C ⋯ \left\{\begin{array}{c} z_0=A x_0+B y_0+C \\ z_1=A x_1+B y_1+C \\ \cdots \end{array}\right. ⎩ ⎨ ⎧z0=Ax0+By0+Cz1=Ax1+By1+C⋯
但由于测量误差的存在,三维点并不都严格符合在一个平面上,所以上述方程无解,需要用最小二乘法来求解上述方程组。
构建方程目标函数:
J ( A , B , C ) = ∑ i = 0 n ( A x i + B y i + C − z i ) 2 J(A, B, C)=\sum_{i=0}^n\left(A x_i+B y_i+C-z_i\right)^2 J(A,B,C)=i=0∑n(Axi+Byi+C−zi)2
求解A,B,C 使得损失函数J最小,采取最小二乘法,分别对A,B,C求偏导,令其偏导数均为0,如下所示:
{ ∂ J / ∂ A = 2 ∗ ∑ i = 0 n ( A x i + B y i + C − z i ) ∗ x i = 0 ∂ J / ∂ B = 2 ∗ ∑ i = 0 n ( A x i + B y i + C − z i ) ∗ y i = 0 ∂ J / ∂ C = 2 ∗ ∑ i = 0 n ( A x i + B y i + C − z i ) = 0 \left\{\begin{array}{c} \partial J / \partial A=2 * \sum_{i=0}^n\left(A x_i+B y_i+C-z_i\right) * x_i=0 \\ \partial J / \partial B=2 * \sum_{i=0}^n\left(A x_i+B y_i+C-z_i\right) * y_i=0 \\ \partial J / \partial C=2 * \sum_{i=0}^n\left(A x_i+B y_i+C-z_i\right)=0 \end{array}\right. ⎩ ⎨ ⎧∂J/∂A=2∗∑i=0n(Axi+Byi+C−zi)∗xi=0∂J/∂B=2∗∑i=0n(Axi+Byi+C−zi)∗yi=0∂J/∂C=2∗∑i=0n(Axi+Byi+C−zi)=0
展开,变形得到:
{ ∑ 2 ( A x i + B y i + C − z i ) x i = 0 ∑ 2 ( A x i + B y i + C − z i ) y i = 0 ∑ 2 ( A x i + B y i + C − z i ) = 0 \left\{\begin{array}{l} \sum 2\left(A x_i+B y_i+C-z_i\right) x_i=0 \\ \sum 2\left(A x_i+B y_i+C-z_i\right) y_i=0 \\ \sum 2\left(A x_i+B y_i+C-z_i\right)=0 \end{array}\right. ⎩ ⎨ ⎧∑2(Axi+Byi+C−zi)xi=0∑2(Axi+Byi+C−zi)yi=0∑2(Axi+Byi+C−zi)=0
其中,A, B, C为变量的线性方程组,写为矩阵形式有:
∣ ∑ x i 2 ∑ x i y i ∑ x i ∑ x i y i ∑ y i 2 ∑ y i ∑ x i ∑ y i n ∣ ⋅ ∣ A B C ∣ = ∣ ∑ z i x i ∑ z i y i ∑ z i ∣ \left|\begin{array}{ccc} \sum \mathrm{x}_{\mathrm{i}}^2 & \sum \mathrm{x}_{\mathrm{i}} \mathrm{y}_{\mathrm{i}} & \sum \mathrm{x}_{\mathrm{i}} \\ \sum \mathrm{x}_{\mathrm{i}} \mathrm{y}_{\mathrm{i}} & \sum \mathrm{y}_{\mathrm{i}}^2 & \sum \mathrm{y}_{\mathrm{i}} \\ \sum \mathrm{x}_{\mathrm{i}} & \sum \mathrm{y}_{\mathrm{i}} & \mathrm{n} \end{array}\right| \cdot\left|\begin{array}{c} \mathrm{A} \\ \mathrm{B} \\ \mathrm{C} \end{array}\right|=\left|\begin{array}{c} \sum \mathrm{z}_{\mathrm{i}} \mathrm{x}_{\mathrm{i}} \\ \sum \mathrm{z}_{\mathrm{i}} \mathrm{y}_{\mathrm{i}} \\ \sum \mathrm{z}_{\mathrm{i}} \end{array}\right| ∑xi2∑xiyi∑xi∑xiyi∑yi2∑yi∑xi∑yin ⋅ ABC = ∑zixi∑ziyi∑zi
构建矩阵方程Ax = b, 其中A为上述方程中左边的参数矩阵方程, x为变量(ABC), b为等式右边的变量。
通过矩阵运算得到 x = A − 1 ∗ b x=\text { A }^{-1} * b x= A −1∗b, 最终得到参数向量[A B C]。
随机给一组点进行测试:(该点为三维坐标系中Oxy平面附近的一组点,坐标的z值在0附近小范围波动,拟合的平面法向量应该近似等于(0,0,1))。
完整代码如下:
#include
#include
#include
void CaculatefitPlane(std::vector points, std::vector &res)
{
// 最小二乘法拟合平面 x = A^-1 * B
// step1 create matrix of A, B, X
cv::Mat A = cv::Mat::zeros(3, 3, CV_64FC1); // Matrix
cv::Mat B = cv::Mat::zeros(3, 1, CV_64FC1); // vector
cv::Mat X = cv::Mat::zeros(3, 1, CV_64FC1); // vector
// step2 input points
double xi = 0;
double xi2 = 0;
double xiyi = 0;
double yi = 0;
double yi2 = 0;
double zi = 0;
double zixi = 0;
double ziyi = 0;
for (int i = 0; i(0, 0) = xi2;
A.at(1, 0) = xiyi;
A.at(2, 0) = xi;
A.at(0, 1) = xiyi;
A.at(1, 1) = yi2;
A.at(2, 1) = yi;
A.at(0, 2) = xi;
A.at(1, 2) = yi;
A.at(2, 2) = points.size();
B.at(0, 0) = zixi;
B.at(1, 0) = ziyi;
B.at(2, 0) = zi;
// step3 calculate plane
// Ax+by+cz=D, c = 1
X = A.inv() * B;
//A
res.push_back(X.at(0, 0));
//B
res.push_back(X.at(1, 0));
//Z的系数为1
res.push_back(1.0);
//C
res.push_back(X.at(2, 0));
return;
}
int main()
{
std::vector points3d;
std::vector planeFun;
points3d.push_back(cv::Point3f(10.1, 20.5, 0.12));
points3d.push_back(cv::Point3f(15.1, 34.5, 0.1));
points3d.push_back(cv::Point3f(13.1, 7.5, -0.05));
points3d.push_back(cv::Point3f(10.1, 25.5, 0.03));
points3d.push_back(cv::Point3f(14.1, 10.5, 0.1));
points3d.push_back(cv::Point3f(16.1, 40.5, 0.2));
points3d.push_back(cv::Point3f(32.1, 10.5, -0.2));
CaculatefitPlane(points3d, planeFun);
for (int i = 0; i < planeFun.size();i++)
{
std::cout << planeFun[i] << std::endl;
}
}
for (int i = 0; i < planeFun.size();i++)
{
std::cout << planeFun[i] << std::endl;
}
}