引言
本来上一篇帖子就已经达到了精度要求,不过经过同学提醒才发现老师的作业要求中有要求考虑归一化。
emmmmm
坐标归一化
进行归一化的必要性和方法参考 《计算机视觉中的多视几何》中的描述:
上面的是从 2D到2D的结论,不过与从3D到2D的结论是相似的:
总结一下,就是在计算之前,要将空间点和平面点都进行归一化,空间点要归一化到距离原点的平均距离为√3,平面点要归一化到距离原点的平均距离为√2。在计算完过后在进行去归一化。其他的计算过程和上一篇帖子讲的完全相同。
如何将点位归一化到这种状态可以参考这篇帖子:对三维点集的归一化变换
MATLAB代码
%%
clc;clear;close all;
%% DLT 算法需要用到 6 对点 ,前三列是 X Y Z 表示空间坐标 ,后两列 x y 表示图像坐标,下面是测试数据
Points = [ 99.1517085791261, -0.892099762666786, 3.06159510601795, 163.209290388816, 182.199675950568;
100.558334460204, 0.868021368458366, 2.25981241694746, 608.892566967667, 701.375883991155;
100.137647321744, -0.0612187178835884, 2.02380413900248, 449.262241640596, 377.206100107244;
99.6742452887978, -0.675635383613515, 3.58856908136781, 316.188427147902, 257.277246206713;
99.6224300840896, 0.0570662710124255, 2.33129745899956, 267.575494465364, 430.054554508315;
100.203963882803, -0.474057430919711, 3.30815819695356, 439.688462560083, 287.229433675101]';
P3d = Points(1:3,:);
P2d = Points(4:5,:);
% 转为齐次坐标
P3d_ = cat(1,P3d,[1,1,1,1,1,1]);
P2d_ = cat(1,P2d,[1,1,1,1,1,1]);
% 求归一化的缩放矩阵
T1 = normalize3d(P3d);
T2 = normalize2d(P2d);
nP3d_ = T1*P3d_;
nP2d_ = T2*P2d_;
% 赋值 有6 个点,所以C的大小为 12×12,
C = zeros(12,12);
for i = 1:6
X = nP3d_(1,i);Y = nP3d_(2,i);Z = nP3d_(3,i);x = nP2d_(1,i);y = nP2d_(2,i);
C(i*2-1,:) = [X,Y,Z,1,0,0,0,0,-x*X,-x*Y,-x*Z,-x];
C(i*2,:) = [0,0,0,0,X,Y,Z,1,-y*X,-y*Y,-y*Z,-y];
end
[U,S,V] = svd(C);
L = V(:,end);
P_ = reshape(L,[4,3])';
% 去归一化
P = inv(T2)*P_*T1;
P = P/P(3,3);
%% 验证
Point3d = ones(4,6);
Point3d(1:3,:) = Points(1:3,:);
Point2d = Points(4:5,:);
Res = P*Point3d;
Res2d = Res(1:2,:)./Res(3,:);
disp(Point2d)
disp(Res2d)
fprintf("重投影误差:\n");
diff = Point2d - Res2d;
disp(diff);
function [T] = normalize3d(w)
% 变换目标:中心在原点且到原点的平均距离为sqrt(3)
if size(w,1)<size(w,2)
w = w';
end
t = -mean(w(:,1:3),1);
w = w(:,1:3) + repmat(t,[size(w,1),1]);
dis = mean(sqrt(w(:,1).*w(:,1)+w(:,2).*w(:,2)+w(:,3).*w(:,3)));
v = sqrt(3)/dis;
t = t'*v;
T = [eye(3)*v,t];
T = [T;0,0,0,1];
end
function [T] = normalize2d(w)
% 变换目标:中心在原点且到原点的平均距离为sqrt(2)
if size(w,1)<size(w,2)
w = w';
end
t = -mean(w(:,1:2),1);
w = w(:,1:2) + repmat(t,[size(w,1),1]);
dis = mean(sqrt(w(:,1).*w(:,1)+w(:,2).*w(:,2)));
v = sqrt(2)/dis;
t = t'*v;
T = [eye(2)*v,t];
T = [T;0,0,1];
end
运行结果
结语
可以看出,相较于没有进行归一化(上一篇帖子)的结果,精度确实得到了提高。
至此,整个DLT便算是彻底完成了!