立体匹配中的NCC，SAD，SSD算法
Normalized Cross correlation (NCC)
NCC(u,v) = [(wl - w)/(|wl - w|)]*[(wr - w)/(|wr - w|)] 选择最大值
Sum of Squared Defferences (SSD)
SSD(u,v) = Sum{[Left(u,v) - Right(u,v)] * [Left(u,v) - Right(u,v)]} 选择最大值
Sum of Absolute Defferences (SAD)
SAD(u,v) = Sum{|Left(u,v) - Right(u,v)|} 选择最小值
先说说SAD算法的基本流程：
1.构造一个小窗口，类似与卷积核。
2.用窗口覆盖左边的图像，选择出窗口覆盖区域内的所有像素点。
3.同样用窗口覆盖右边的图像并选择出覆盖区域的像素点。
4.左边覆盖区域减去右边覆盖区域，并求出所有像素点差的绝对值的和。
5.移动右边图像的窗口，重复3，4的动作。（这里有个搜索范围，超过这个范围跳出）
6.找到这个范围内SAD值最小的窗口，即找到了左边图像的最佳匹配的像素块。
OpenCV代码示范SAD:
代码
1 IplImage* generateDisparityImage(IplImage* greyLeftImg32,
2 IplImage* greyRightImg32,
3 int windowSize,int DSR){
4
5 int offset=floor((double)windowSize/2);
6 int height=greyLeftImg32->height;
7 int width=greyLeftImg32->width;
8 double* localSAD=new double[DSR];//DSR即搜索范围
9
10 int x=0, y=0,d=0,m=0;
11 int N=windowSize;
12
13 IplImage* winImg=cvCreateImage(cvSize(N,N),32,1);//mySubImage(greyLeftImg32,cvRect(0,0,N,N));
14
15 IplImage* disparity=cvCreateImage(cvSize(width,height),8,1);//or IPL_DEPTH_8U
16 BwImage imgA(disparity);
17
18 for (y=0;y<height;y++){
19 for (x=0;x<width;x++){
20 imgA[y][x]=0;
21 }
22 }
23
24 CvScalar sum;
25 //CvScalar s2;
26 for (y=0;y<height-N;y++){ //height-N
27 for (x=0;x<width-N;x++){//width-N
28 cvSetImageROI(greyLeftImg32, cvRect(x,y,N,N));
29 d=0;
30 //initialise localSAD
31 for (m=0;m<DSR;m++){localSAD[m]=0;}
32
33 //start matching
34 do{
35 if (x-d>=0){
36 cvSetImageROI(greyRightImg32, cvRect(x-d,y,N,N));
37 }else{
38 break;
39 }
40
41 cvAbsDiff(greyLeftImg32,greyRightImg32,winImg);//absolute difference
42 sum=cvSum(winImg);//sum
43 localSAD[d]=sum.val[0];//0 means single channel
44
45 cvResetImageROI(greyRightImg32);
46 d++;
47 }while(d<=DSR);
48
49 //to find the best d and store
50 imgA[y+offset][x+offset]=getMaxMin(localSAD,DSR,0)*16; //0 means return minimum index
51 cvResetImageROI(greyLeftImg32);
52 }//x
53 if (y%10==0)cout<<"row="<<y<<" of "<<height<<endl;
54 }//y
55
56 cvReleaseImage(&winImg);
57 //cvReleaseImage(&rightWinImg);
58
59 return disparity;
60
61 }
2 IplImage* greyRightImg32,
3 int windowSize,int DSR){
4
5 int offset=floor((double)windowSize/2);
6 int height=greyLeftImg32->height;
7 int width=greyLeftImg32->width;
8 double* localSAD=new double[DSR];//DSR即搜索范围
9
10 int x=0, y=0,d=0,m=0;
11 int N=windowSize;
12
13 IplImage* winImg=cvCreateImage(cvSize(N,N),32,1);//mySubImage(greyLeftImg32,cvRect(0,0,N,N));
14
15 IplImage* disparity=cvCreateImage(cvSize(width,height),8,1);//or IPL_DEPTH_8U
16 BwImage imgA(disparity);
17
18 for (y=0;y<height;y++){
19 for (x=0;x<width;x++){
20 imgA[y][x]=0;
21 }
22 }
23
24 CvScalar sum;
25 //CvScalar s2;
26 for (y=0;y<height-N;y++){ //height-N
27 for (x=0;x<width-N;x++){//width-N
28 cvSetImageROI(greyLeftImg32, cvRect(x,y,N,N));
29 d=0;
30 //initialise localSAD
31 for (m=0;m<DSR;m++){localSAD[m]=0;}
32
33 //start matching
34 do{
35 if (x-d>=0){
36 cvSetImageROI(greyRightImg32, cvRect(x-d,y,N,N));
37 }else{
38 break;
39 }
40
41 cvAbsDiff(greyLeftImg32,greyRightImg32,winImg);//absolute difference
42 sum=cvSum(winImg);//sum
43 localSAD[d]=sum.val[0];//0 means single channel
44
45 cvResetImageROI(greyRightImg32);
46 d++;
47 }while(d<=DSR);
48
49 //to find the best d and store
50 imgA[y+offset][x+offset]=getMaxMin(localSAD,DSR,0)*16; //0 means return minimum index
51 cvResetImageROI(greyLeftImg32);
52 }//x
53 if (y%10==0)cout<<"row="<<y<<" of "<<height<<endl;
54 }//y
55
56 cvReleaseImage(&winImg);
57 //cvReleaseImage(&rightWinImg);
58
59 return disparity;
60
61 }
