Python：对3D点拟合3D椭球（扁平/长形）

Question

亲爱的stackoverflow用户们，

我面临以下问题：我想在我的Python脚本中将3D椭球拟合到3D数据点。

起始数据是一组x、y和z坐标（笛卡尔坐标）。我想要得到的是最佳拟合椭球的定义方程中的a和c，这个椭球是由3D数据点的凸包定义的。

方程在适当旋转和平移后的坐标系中为：

所以我理想情况下想要做的任务是：

1. 找出3D数据点的凸包

2. 将最佳拟合椭球拟合到凸包并得到a和c

您是否知道是否有一些库可以用最少的代码行在Python中完成这些操作？还是说我必须使用我的有限数学知识（实际上几乎为零，当涉及到找到最佳拟合椭球时）明确编写每一个步骤？

- Glxblt76

---

如果“正确旋转和平移的坐标系”已知，则只有两个未知数，并且可以将其重写为易于线拟合的问题A（x² + y²）/ z² + B = 1 / z²。但是您知道这个系统吗？ - user1196549

你如何称呼“最佳拟合”？你允许凸壳的某些顶点在“最佳拟合”之外吗？ - user1196549

@YvesDaoust 我终于找到了解决方案。最佳拟合是减少点与椭球之间平方距离的那个。 - Glxblt76

我很好奇你为什么要把寻找凸包作为第一步。如果你的初始点存在噪声，凸包可能会产生扭曲的椭球拟合。 - EA304GT

Answer 1

好的，我通过将Scipy的凸包算法与此网站上找到的一些Python函数相结合，找到了我的解决方案。

假设您得到了名为x、y和z的numpy向量的X坐标、Y坐标和Z坐标。这对我有效：

from   scipy.spatial            
import ConvexHull, convex_hull_plot_2d
import numpy as np
from   numpy.linalg import eig, inv

def ls_ellipsoid(xx,yy,zz):                                  
    #finds best fit ellipsoid. Found at http://www.juddzone.com/ALGORITHMS/least_squares_3D_ellipsoid.html
    #least squares fit to a 3D-ellipsoid
    #  Ax^2 + By^2 + Cz^2 +  Dxy +  Exz +  Fyz +  Gx +  Hy +  Iz  = 1
    #
    # Note that sometimes it is expressed as a solution to
    #  Ax^2 + By^2 + Cz^2 + 2Dxy + 2Exz + 2Fyz + 2Gx + 2Hy + 2Iz  = 1
    # where the last six terms have a factor of 2 in them
    # This is in anticipation of forming a matrix with the polynomial coefficients.
    # Those terms with factors of 2 are all off diagonal elements.  These contribute
    # two terms when multiplied out (symmetric) so would need to be divided by two
    
    # change xx from vector of length N to Nx1 matrix so we can use hstack
    x = xx[:,np.newaxis]
    y = yy[:,np.newaxis]
    z = zz[:,np.newaxis]
    
    #  Ax^2 + By^2 + Cz^2 +  Dxy +  Exz +  Fyz +  Gx +  Hy +  Iz = 1
    J = np.hstack((x*x,y*y,z*z,x*y,x*z,y*z, x, y, z))
    K = np.ones_like(x) #column of ones
    
    #np.hstack performs a loop over all samples and creates
    #a row in J for each x,y,z sample:
    # J[ix,0] = x[ix]*x[ix]
    # J[ix,1] = y[ix]*y[ix]
    # etc.
    
    JT=J.transpose()
    JTJ = np.dot(JT,J)
    InvJTJ=np.linalg.inv(JTJ);
    ABC= np.dot(InvJTJ, np.dot(JT,K))

    # Rearrange, move the 1 to the other side
    #  Ax^2 + By^2 + Cz^2 +  Dxy +  Exz +  Fyz +  Gx +  Hy +  Iz - 1 = 0
    #    or
    #  Ax^2 + By^2 + Cz^2 +  Dxy +  Exz +  Fyz +  Gx +  Hy +  Iz + J = 0
    #  where J = -1
    eansa=np.append(ABC,-1)

    return (eansa)

def polyToParams3D(vec,printMe):                             
    #gets 3D parameters of an ellipsoid. Found at http://www.juddzone.com/ALGORITHMS/least_squares_3D_ellipsoid.html
    # convert the polynomial form of the 3D-ellipsoid to parameters
    # center, axes, and transformation matrix
    # vec is the vector whose elements are the polynomial
    # coefficients A..J
    # returns (center, axes, rotation matrix)
    
    #Algebraic form: X.T * Amat * X --> polynomial form
    
    if printMe: print('\npolynomial\n',vec)
    
    Amat=np.array(
    [
    [ vec[0],     vec[3]/2.0, vec[4]/2.0, vec[6]/2.0 ],
    [ vec[3]/2.0, vec[1],     vec[5]/2.0, vec[7]/2.0 ],
    [ vec[4]/2.0, vec[5]/2.0, vec[2],     vec[8]/2.0 ],
    [ vec[6]/2.0, vec[7]/2.0, vec[8]/2.0, vec[9]     ]
    ])
    
    if printMe: print('\nAlgebraic form of polynomial\n',Amat)
    
    #See B.Bartoni, Preprint SMU-HEP-10-14 Multi-dimensional Ellipsoidal Fitting
    # equation 20 for the following method for finding the center
    A3=Amat[0:3,0:3]
    A3inv=inv(A3)
    ofs=vec[6:9]/2.0
    center=-np.dot(A3inv,ofs)
    if printMe: print('\nCenter at:',center)
    
    # Center the ellipsoid at the origin
    Tofs=np.eye(4)
    Tofs[3,0:3]=center
    R = np.dot(Tofs,np.dot(Amat,Tofs.T))
    if printMe: print('\nAlgebraic form translated to center\n',R,'\n')
    
    R3=R[0:3,0:3]
    R3test=R3/R3[0,0]
    # print('normed \n',R3test)
    s1=-R[3, 3]
    R3S=R3/s1
    (el,ec)=eig(R3S)
    
    recip=1.0/np.abs(el)
    axes=np.sqrt(recip)
    if printMe: print('\nAxes are\n',axes  ,'\n')
    
    inve=inv(ec) #inverse is actually the transpose here
    if printMe: print('\nRotation matrix\n',inve)
    return (center,axes,inve)


#let us assume some definition of x, y and z

#get convex hull
surface  = np.stack((conf.x,conf.y,conf.z), axis=-1)
hullV    = ConvexHull(surface)
lH       = len(hullV.vertices)
hull     = np.zeros((lH,3))
for i in range(len(hullV.vertices)):
    hull[i] = surface[hullV.vertices[i]]
hull     = np.transpose(hull)         
            
#fit ellipsoid on convex hull
eansa            = ls_ellipsoid(hull[0],hull[1],hull[2]) #get ellipsoid polynomial coefficients
print("coefficients:"  , eansa)
center,axes,inve = polyToParams3D(eansa,False)   #get ellipsoid 3D parameters
print("center:"        , center)
print("axes:"          , axes)
print("rotationMatrix:", inve)

- Glxblt76