逆透视变换详解及代码实现（二）

根据逆透视变换详解及代码实现(一)的原理

下面我用车上拍摄的车道图像，采用逆透视变换得到的图像，给出代码前我们先看下处理结果。

首先是原始图像：

下图为逆透视变换图像：

下面说具体的实现吧！！

一、参数设置：

1、需要知道相机的内部参数（这个具体步骤可以找相关文档，这里就不具体展开说）。

我们这里假设已经获取内部参数：

相机焦距，相机光学中心，相机高度，相机的俯仰角，相机的偏航角，相机拍摄出的图像尺寸。

参数说明：其中偏航角 和 俯仰角 就是在（一）中所说的世界坐标经过旋转矩阵得到相应的相机坐标。而偏航角和俯仰角将决定这个旋转矩阵。而相机焦距 和相机光学中心 是可以从相机标定后得出，相机高度需要自己测量。

而图像尺寸，是拍出图像的尺寸。

2、设定逆透视变换的参数：

逆透视图像的尺寸，需要进行逆透视变换的区域，逆透视变换的差值算法。

逆透视变换的区域：原始图像中需要变换的区域（当然这个区域不能超过消失点区域，后面会说到）

逆透视图像的尺寸：就是需要将逆透视变换区域映射到这个逆透视图像上。

差值算法：因为需要映射，所以某些数值需要估计出，这里用双线性差值。

二、

根据相机的内部参数计算消失点：

因为图像是二维的，所以消失点是是一个二维坐标。

code:

function [ vp ] = GetVanishingPoint( cameraInfo )
%GetVanishingPoint Summary of this function goes here
%   Detailed explanation goes herevpp = [ sin(cameraInfo.yaw*pi/180)/cos(cameraInfo.pitch*pi/180);cos(cameraInfo.yaw*pi/180)/cos(cameraInfo.pitch*pi/180);0];tyawp = [cos(cameraInfo.yaw*pi/180), -sin(cameraInfo.yaw*pi/180), 0;         sin(cameraInfo.yaw*pi/180), cos(cameraInfo.yaw*pi/180), 0;0, 0, 1];tpitchp = [1, 0, 0;0, -sin(cameraInfo.pitch*pi/180), -cos(cameraInfo.pitch*pi/180);0, cos(cameraInfo.pitch*pi/180), -sin(cameraInfo.pitch*pi/180)];transform = tyawp*tpitchp;t1p = [cameraInfo.focalLengthX, 0, cameraInfo.opticalCenterX;0, cameraInfo.focalLengthY, cameraInfo.opticalCenterY;0, 0, 1];
transform = t1p*transform;vp = transform*vpp;end

三、利用消失点可以得到uv平面中的图像范围和对应的xy平面的范围

code:

uvLimitsp = [ vp.x,         ipmInfo.ipmRight, ipmInfo.ipmLeft,  vp.x;ipmInfo.ipmTop, ipmInfo.ipmTop, ipmInfo.ipmTop,   ipmInfo.ipmBottom];xyLimits = TransformImage2Ground(uvLimitsp,cameraInfo);

function [ xyLimits ] = TransformImage2Ground( uvLimits,cameraInfo )
%TransformImage2Ground Summary of this function goes here
%   Detailed explanation goes here
[row , col ] = size(uvLimits);
inPoints4 = zeros(row+2,col);
inPoints4(1:2,:) = uvLimits;
inPoints4(3,:) = 1;
inPoints3 = inPoints4(1:3,:);c1 = cos(cameraInfo.pitch*pi/180);
s1 = sin(cameraInfo.pitch*pi/180);
c2 = cos(cameraInfo.yaw*pi/180);
s2 = sin(cameraInfo.yaw*pi/180);matp= [-cameraInfo.cameraHeight*c2/cameraInfo.focalLengthX,cameraInfo.cameraHeight*s1*s2/cameraInfo.focalLengthY,(cameraInfo.cameraHeight*c2*cameraInfo.opticalCenterX/cameraInfo.focalLengthX)- (cameraInfo.cameraHeight *s1*s2* cameraInfo.opticalCenterY/ cameraInfo.focalLengthY) - cameraInfo.cameraHeight *c1*s2;cameraInfo.cameraHeight *s2/cameraInfo.focalLengthX, ...cameraInfo.cameraHeight *s1*c2/cameraInfo.focalLengthY, ...(-cameraInfo.cameraHeight *s2* cameraInfo.opticalCenterX ..../cameraInfo.focalLengthX)-(cameraInfo.cameraHeight *s1*c2* ....cameraInfo.opticalCenterY /cameraInfo.focalLengthY) -  ...cameraInfo.cameraHeight *c1*c2;0, cameraInfo.cameraHeight *c1/cameraInfo.focalLengthY, (-cameraInfo.cameraHeight *c1* cameraInfo.opticalCenterY/cameraInfo.focalLengthY)+cameraInfo.cameraHeight*s1;0, -c1 /cameraInfo.focalLengthY,(c1* cameraInfo.opticalCenterY /cameraInfo.focalLengthY) - s1];inPoints4 = matp*inPoints3;
inPointsr4 = inPoints4(4,:);
div = inPointsr4;
inPoints4(1,:) = inPoints4(1,:)./div;
inPoints4(2,:) = inPoints4(2,:)./div;
inPoints2 = inPoints4(1:2,:);
xyLimits = inPoints2;end

四、根据得到的范围计算xy平面的一一对应的映射

row1 = xyLimits(1,:);
row2 = xyLimits(2,:);
xfMin = min(row1); xfMax = max(row1);
yfMin = min(row2); yfMax = max(row2);[outRow outCol] = size(outImage);
stepRow = (yfMax - yfMin)/outRow;
stepCol = (xfMax - xfMin)/outCol;
xyGrid = zeros(2,outRow*outCol);
y = yfMax-0.5*stepRow;for i = 1:outRowx = xfMin+0.5*stepCol;for j = 1:outColxyGrid(1,(i-1)*outCol+j) = x;xyGrid(2,(i-1)*outCol+j) = y;x = x + stepCol;endy = y - stepRow;
end

五、将xy平面的映射转换到uv平面，并画出这个映射

%TransformGround2Image
uvGrid = TransformGround2Image(xyGrid,cameraInfo);
% Image mean 
means = mean(R(:))/255;
RR = double(R)/255;
for i=1:outRowfor j = 1:outCol;ui = uvGrid(1,(i-1)*outCol+j);vi = uvGrid(2,(i-1)*outCol+j);if (ui<ipmInfo.ipmLeft || ui>ipmInfo.ipmRight || vi<ipmInfo.ipmTop || vi>ipmInfo.ipmBottom) outImage(i,j) = means;elsex1 = int32(ui); x2 = int32(ui+1);y1 = int32(vi); y2 = int32(vi+1);x = ui-double(x1) ;  y = vi-double(y1);val = double(RR(y1,x1))*(1-x)*(1-y)+double(RR(y1,x2))*x*(1-y)+double(RR(y2,x1))*(1-x)*y+double(RR(y2,x2))*x*y;outImage(i,j) = val;endend
end