双目立体匹配方法：BM、SGBM、GC算法的实现及性能对比(附代码)

在双目立体匹配中，有基于SAD算法的BM、SGBM、GC立体匹配技术，但是在OpenCv3.0以后，GC算法就从legacy中去除掉了。为了查看三种算法的匹配效果及运算性能，我在Windows10中安装了OpenCv2.4.9，并以VsCode为IDE进行程序编写。

一、VsCode中集成OpenCv2.4.9

VsCode中集成OpenCv2.4.9，我主要是参照了这个文章的方法，根据这位博主的方法，我完成了环境的配置，由于需要相应的legacy库，所以我的tasks.json如下所示

{
    // See https://go.microsoft.com/fwlink/?LinkId=733558
    // for the documentation about the tasks.json format
    "version": "2.0.0",
    "command": "g++",
    "args": [
        "-g", 
        "-std=c++11", 
        "${file}", 
        "-o", 
        "${fileBasenameNoExtension}.exe",  
        "-I", "F:\\opencv\\build\\include",
        "-I", "F:\\opencv\\build\\include\\opencv2",
        "-I", "F:\\opencv\\build\\include\\opencv",
        "-L", "F:\\opencv\\build\\x64\\MinGW\\lib",
        "-l", "opencv_core249",
        "-l", "opencv_imgproc249",
        "-l", "opencv_video249",
        "-l", "opencv_ml249",
        "-l", "opencv_highgui249",
        "-l", "opencv_objdetect249",
        "-l", "opencv_flann249",
        "-l", "opencv_photo249",
        "-l", "opencv_legacy249",
        "-l", "opencv_calib3d249"
    ],// 编译命令参数
    "problemMatcher":{
        "owner": "cpp",
        "fileLocation":[
            "relative",
            "${workspaceFolder}"
        ],
        "pattern":[
            {
                "regexp": "^([^\\\\s].*)\\\\((\\\\d+,\\\\d+)\\\\):\\\\s*(.*)$",
                "file": 1,
                "location": 2,
                "message": 3
            }
        ]
    },
    "group": {
        "kind": "build",
        "isDefault": true
    }
  }
  

二、程序的实现

程序实现的文章参考了OpenCV三种立体匹配求视差图算法总结这篇博客，同时也有对官网例程的参考。程序的完整代码如下：

#include <ctime> 
#include <math.h>  
#include <iostream>

#include "opencv2/calib3d/calib3d.hpp"    
#include "opencv2/imgproc/imgproc.hpp"    
#include "opencv2/highgui/highgui.hpp"    
#include "opencv2/contrib/contrib.hpp"   
#include "opencv2/legacy/legacy.hpp" 
#include <opencv2/nonfree/features2d.hpp>
#include <opencv2/features2d/features2d.hpp>

using namespace std;
using namespace cv;

enum { STEREO_BM=0, STEREO_SGBM=1, STEREO_HH=2, STEREO_VAR=3 };
int alg = STEREO_SGBM;

int main()
{
    time_t begin,end;
    double ret;
    int SADWindowSize = 0, numberOfDisparities = 0;

    IplImage * img1 = cvLoadImage("D:/projects/images/Cones_l.png",0);
    IplImage * img2 = cvLoadImage("D:/projects/images/Cones_r.png",0);

    Mat left = imread("D:/projects/images/Cones_l.png",0);
    Mat right = imread("D:/projects/images/Cones_r.png",0);
    Size img_size = left.size();

    CvMat* displeft=cvCreateMat(img1->height,img1->width,CV_16S);
    CvMat* dispright=cvCreateMat(img2->height,img2->width,CV_16S);
    CvMat* disp=cvCreateMat(img1->height,img1->width,CV_8U);

    //BM算法
    CvStereoBMState *BMState = cvCreateStereoBMState();
    numberOfDisparities = numberOfDisparities > 0 ? numberOfDisparities : ((img_size.width/8) + 15) & -16;
    BMState->SADWindowSize = SADWindowSize > 0 ? SADWindowSize : 9;
    BMState->minDisparity = 0;
    BMState->numberOfDisparities = numberOfDisparities;
    BMState->textureThreshold = 10;
    BMState->uniquenessRatio = 15;
    BMState->speckleWindowSize = 100;
    BMState->speckleRange = 32;
    BMState->disp12MaxDiff = 1;
    begin=clock();
    cvFindStereoCorrespondenceBM( img1, img2, displeft,BMState);
    end=clock();
    ret=double(end-begin)/CLOCKS_PER_SEC;
    cout<<"BM algorithm runtime:   "<<ret<<" s"<<endl;
    cvNormalize( displeft, disp, 0, 255, CV_MINMAX );
    cvSaveImage("BM_right_disparity.png",disp);

    cvNamedWindow("BM_disparity",0);
    cvShowImage("BM_disparity",disp);


    //SGBM算法
    StereoSGBM sgbm;
	sgbm.preFilterCap = 63;
    sgbm.SADWindowSize = SADWindowSize > 0 ? SADWindowSize : 5;
    numberOfDisparities = numberOfDisparities > 0 ? numberOfDisparities : ((img_size.width/8) + 15) & -16;
	int cn = left.channels();
	sgbm.P1 = 8*cn*sgbm.SADWindowSize*sgbm.SADWindowSize;
	sgbm.P2 = 32*cn*sgbm.SADWindowSize*sgbm.SADWindowSize;
	sgbm.minDisparity = 0;
	sgbm.numberOfDisparities = numberOfDisparities;
	sgbm.uniquenessRatio = 10;
	sgbm.speckleWindowSize = 100;
	sgbm.speckleRange = 32;
	sgbm.disp12MaxDiff = 1;
    sgbm.fullDP = alg == STEREO_HH;
    Mat disp_, disp8;
	//sgbm(left , right , left_disp);
	begin=clock();
	sgbm(left, right, disp_);
    disp_.convertTo(disp8, CV_8U,255/(numberOfDisparities*16.));
    end=clock();
    ret=double(end-begin)/CLOCKS_PER_SEC;
    cout<<"SGBM algorithm runtime:   "<<ret<<" s"<<endl;
    imwrite("SGBM_disparity.png",disp8);
    namedWindow("SGBM_disparity",WINDOW_AUTOSIZE);
    imshow("SGBM_disparity",disp8);

    //GC算法
    numberOfDisparities = numberOfDisparities > 0 ? numberOfDisparities : ((img1->width/8) + 15) & -16;
    CvStereoGCState* GCState=cvCreateStereoGCState(numberOfDisparities,2);   //原始为64 3
    assert(GCState);
    CvMat* gcdispleft=cvCreateMat(img1->height,img1->width,CV_16S);
    CvMat* gcdispright=cvCreateMat(img2->height,img2->width,CV_16S);
    CvMat* gcvdisp=cvCreateMat(img1->height,img1->width,CV_8U);    
    begin=clock();  
    cvFindStereoCorrespondenceGC(img1,img2,gcdispleft,gcdispright,GCState);
    end=clock();
    ret=double(end-begin)/CLOCKS_PER_SEC;
    cout<<"GC algorithm runtime:   "<<ret<<" s"<<endl;
    cvNormalize(gcdispright,gcvdisp,0,255,CV_MINMAX);
    cvSaveImage("GC_disparity.png",gcvdisp);
    cvNamedWindow("GC_disparity",0);
    cvShowImage("GC_disparity",gcvdisp);
    cvWaitKey(0);
    cvReleaseMat(&displeft);
    cvReleaseMat(&dispright);
    cvReleaseMat(&disp);
    cvReleaseMat(&gcdispleft);
    cvReleaseMat(&gcdispright);
    cvReleaseMat(&gcvdisp);    
    return 0;

三、效果对比

我的电脑为联想拯救者Y7000，在上面运行，三种算法的时间分别为BM:8ms,SGBM:33ms,GC:2030ms，效果图如下：
Cones_l.png

Cones_r.png

BM算法

SGBM算法

GC算法