【OpenCV学习】计算两幅图像的重叠区域

问题描述：已知两幅图像Image1和Image2，计算出两幅图像的重叠区域，并在Image1和Image2标识出重叠区域。

算法思想：

若两幅图像存在重叠区域，则进行图像匹配后，会得到一张完整的全景图，因而可以转换成图像匹配问题。

图像匹配问题，可以融合两幅图像，得到全景图，但无法标识出在原图像的重叠区域。

将两幅图像都理解为多边形，则其重叠区域的计算，相当于求多边形的交集。

通过多边形求交，获取重叠区域的点集，然后利用单应矩阵还原在原始图像的点集信息，从而标识出重叠区域。

算法步骤：

1.图像匹配计算，获取单应矩阵。

2.根据单应矩阵，计算图像2的顶点转换后的点集。

3.由图像1的顶点集合和图像2的转换点集，计算多边形交集。

4.根据单应矩阵的逆，计算多边形的交集在图像2中的原始点集。

代码实现如下所示:

  1 bool ImageOverlap(cv::Mat &img1,cv::Mat &img2,std::vector<cv::Point> &vPtsImg1,std::vector<cv::Point> &vPtsImg2)

  2 {

  3     cv::Mat g1(img1,Rect(0,0,img1.cols,img1.rows));

  4     cv::Mat g2(img2,Rect(0,0,img2.cols,img2.rows));

  5 

  6     cv::cvtColor(g1,g1,CV_BGR2GRAY);

  7     cv::cvtColor(g2,g2,CV_BGR2GRAY);

  8 

  9     std::vector<cv::KeyPoint> keypoints_roi, keypoints_img;  /* keypoints found using SIFT */

 10     cv::Mat descriptor_roi, descriptor_img;             /* Descriptors for SIFT */

 11     cv::FlannBasedMatcher matcher;                      /* FLANN based matcher to match keypoints */

 12     std::vector<cv::DMatch> matches, good_matches;

 13     cv::SIFT sift;

 14     int i, dist=80;

 15 

 16     sift(g1, Mat(), keypoints_roi, descriptor_roi);      /* get keypoints of ROI image */

 17     sift(g2, Mat(), keypoints_img, descriptor_img);         /* get keypoints of the image */

 18     matcher.match(descriptor_roi, descriptor_img, matches);

 19 

 20     double max_dist = 0; double min_dist = 1000;

 21 

 22     //-- Quick calculation of max and min distances between keypoints

 23     for( int i = 0; i < descriptor_roi.rows; i++ )

 24     { 

 25         double dist = matches[i].distance;

 26         if( dist < min_dist ) min_dist = dist;

 27         if( dist > max_dist ) max_dist = dist;

 28     }

 29 

 30     for (i=0; i < descriptor_roi.rows; i++)

 31     {

 32         if (matches[i].distance < 3*min_dist)

 33         {

 34             good_matches.push_back(matches[i]);

 35         }

 36     }

 37 

 38     //printf("%ld no. of matched keypoints in right image\n", good_matches.size());

 39     /* Draw matched keypoints */

 40 

 41     //Mat img_matches;

 42     //drawMatches(img1, keypoints_roi, img2, keypoints_img, 

 43     //    good_matches, img_matches, Scalar::all(-1), 

 44     //    Scalar::all(-1), vector<char>(), 

 45     //    DrawMatchesFlags::NOT_DRAW_SINGLE_POINTS);

 46     //imshow("matches",img_matches);

 47 

 48     vector<Point2f> keypoints1, keypoints2; 

 49     for (i=0; i<good_matches.size(); i++)

 50     {

 51         keypoints1.push_back(keypoints_img[good_matches[i].trainIdx].pt);

 52         keypoints2.push_back(keypoints_roi[good_matches[i].queryIdx].pt);

 53     }

 54     //计算单应矩阵

 55     Mat H = findHomography( keypoints1, keypoints2, CV_RANSAC );

 56 

 57     //show stitchImage

 58     // cv::Mat stitchedImage;

 59     // int mRows = img2.rows;

 60     // if (img1.rows> img2.rows)

 61     // {

 62         // mRows = img1.rows;

 63     // }

 64     // stitchedImage = Mat::zeros(img2.cols+img1.cols, mRows, CV_8UC3);

 65     // warpPerspective(img2,stitchedImage,H,Size(img2.cols+img1.cols,mRows));

 66     // Mat half(stitchedImage,Rect(0,0,img1.cols,img1.rows));

 67     // img1.copyTo(half);

 68     // imshow("stitchedImage",stitchedImage);

 69 

 70     std::vector<cv::Point> vSrcPtsImg1;

 71     std::vector<cv::Point> vSrcPtsImg2;

 72 

 73     vSrcPtsImg1.push_back(cv::Point(0,0));

 74     vSrcPtsImg1.push_back(cv::Point(0,img1.rows));

 75     vSrcPtsImg1.push_back(cv::Point(img1.cols,img1.rows));

 76     vSrcPtsImg1.push_back(cv::Point(img1.cols,0));

 77 

 78     vSrcPtsImg2.push_back(cv::Point(0,0));

 79     vSrcPtsImg2.push_back(cv::Point(0,img2.rows));

 80     vSrcPtsImg2.push_back(cv::Point(img2.cols,img2.rows));

 81     vSrcPtsImg2.push_back(cv::Point(img2.cols,0));

 82 

 83     //计算图像2在图像1中对应坐标信息

 84     std::vector<cv::Point> vWarpPtsImg2;

 85     for(int i = 0;i < vSrcPtsImg2.size();i++ )

 86     {

 87         cv::Mat srcMat = Mat::zeros(3,1,CV_64FC1);

 88         srcMat.at<double>(0,0) = vSrcPtsImg2[i].x;

 89         srcMat.at<double>(1,0) = vSrcPtsImg2[i].y;

 90         srcMat.at<double>(2,0) = 1.0;

 91         

 92         cv::Mat warpMat = H * srcMat;

 93         cv::Point warpPt;

 94         warpPt.x = cvRound(warpMat.at<double>(0,0)/warpMat.at<double>(2,0));

 95         warpPt.y = cvRound(warpMat.at<double>(1,0)/warpMat.at<double>(2,0));

 96 

 97         vWarpPtsImg2.push_back(warpPt);

 98     }

 99     //计算图像1和转换后的图像2的交点

100     if(!PolygonClip(vSrcPtsImg1,vWarpPtsImg2,vPtsImg1))

101         return false;

102 

103     for (int i = 0;i < vPtsImg1.size();i++)

104     {

105         cv::Mat srcMat = Mat::zeros(3,1,CV_64FC1);

106         srcMat.at<double>(0,0) = vPtsImg1[i].x;

107         srcMat.at<double>(1,0) = vPtsImg1[i].y;

108         srcMat.at<double>(2,0) = 1.0;

109 

110         cv::Mat warpMat = H.inv() * srcMat;

111         cv::Point warpPt;

112         warpPt.x = cvRound(warpMat.at<double>(0,0)/warpMat.at<double>(2,0));

113         warpPt.y = cvRound(warpMat.at<double>(1,0)/warpMat.at<double>(2,0));

114         vPtsImg2.push_back(warpPt);

115     }

116     return true;

117 }

其中，多边形求交集可参考：http://www.cnblogs.com/dwdxdy/p/3232110.html

最终，程序运行的示意图如下：