Surf特徵匹配點提純

本文將介紹Surf特徵匹配得到的特徵匹配點提純方法

      surf匹配代碼：

cv::Ptr<Feature2D> p_ex1 = xfeatures2d::SURF::create();//生成對應描述符
			p_ex1->detect(thisFHalf_1, thisFrameKeypoints);
			p_ex1->detect(nextFHalf_2, nextFrameKeypoints);

			p_ex1->compute(thisFHalf_1, thisFrameKeypoints, thisFrameDescriptor);
			p_ex1->compute(nextFHalf_2, nextFrameKeypoints, nextFrameDescriptor);

std::vector< cv::DMatch > matches_1;
			//std::vector< cv::DMatch > matches_2;
			std::vector<cv::DMatch> good_matches;
			std::vector<cv::DMatch> good_matches2;
			BFMatcher matcher;
			//FlannBasedMatcher matcher;
			//matcher.knnMatch()
			matcher.match(thisFrameDescriptor, nextFrameDescriptor, matches_1);

獲得匹配點對matches_1,但是surf獲得的匹配點對中存在相當多的誤匹配，如何剔除誤匹配，提取精度較高的匹配點對。

（一）通過距離判斷來提純，這也是常用的方法      

//消除錯配點
			double min_dist = 100;
			double max_dist = 0;
			for (int i = 0; i < matches_1.size(); i++)
			{
				double dist = matches_1[i].distance;
				if (dist < min_dist)
				{
					min_dist = dist;
				}
				if (dist > max_dist)
				{
					max_dist = dist;
				}
			}
			//std::vector<cv::DMatch> good_matches;

			for (int i = 0; i < matches_1.size(); i++)
			{
				if (matches_1[i].distance < (3 * min_dist))
				{
					good_matches.push_back(matches_1[i]);
					//surfgoodnum++;
				}
			}

通常設置3*min_dist，距離越小精度也高，但數量也少。

（二）正反匹配提純

cv::Mat img_match;
int num = 0;
for (int a = 0; a < matches_1.size(); a++)
{
	for (int b = 0; b < matches_2.size();b++)
	{
            if (thisFrameKeypoints[matches_1[a].queryIdx].pt == thisFrameKeypoints[matches_2[b].trainIdx].pt && nextFrameKeypoints[matches_1[a].trainIdx].pt == nextFrameKeypoints[matches_2[b].queryIdx].pt)
	    {
		if (fabs(thisFrameKeypoints[matches_1[a].queryIdx].pt.x - thisFrameKeypoints[matches_1[a].trainIdx].pt.x)<20 && fabs(thisFrameKeypoints[matches_1[a].queryIdx].pt.y - thisFrameKeypoints[matches_1[a].trainIdx].pt.y)<20)
		{
			good_matches.push_back(matches_1[a]);
		}
						
	 }
     }
}

如果前後都能匹配則認爲該點對爲匹配點對。（效果不佳，只能去除部分）

（三）單應消除錯誤匹配點

//單應去除匹配錯點
<span style="white-space:pre">	</span>std::vector<cv::Point2f> srcPoints_1, dstPoints_1;
<span style="white-space:pre">	</span>for (std::vector< cv::DMatch >::iterator it = matches_1.begin(); it != matches_1.end(); it++)
<span style="white-space:pre">	</span>{
<span style="white-space:pre">	</span>    srcPoints_1.push_back(thisFrameKeypoints[it->queryIdx].pt);//ThisFrame
<span style="white-space:pre">	</span>    dstPoints_1.push_back(nextFrameKeypoints[it->trainIdx].pt);//NextFrame
<span style="white-space:pre">	</span>}
<span style="white-space:pre">	</span>cv::Mat H_1;
<span style="white-space:pre">	</span>float reprojectionThreshold = 0.1;

        std::vector<cv::DMatch> inliers;
<span style="white-space:pre">	</span>std::vector<unsigned char> inliersMask(srcPoints_1.size());
<span style="white-space:pre">	</span>H_1 = findHomography(
<span style="white-space:pre">	</span>      srcPoints_1, dstPoints_1,
<span style="white-space:pre">	</span>      CV_FM_RANSAC, reprojectionThreshold,
<span style="white-space:pre">	</span>      inliersMask);
<span style="white-space:pre">	</span>      for (size_t i = 0; i < inliersMask.size(); i++)
<span style="white-space:pre">		</span>{
<span style="white-space:pre">		</span>   if (inliersMask[i])
<span style="white-space:pre">			</span>inliers.push_back(matches_1[i]);
<span style="white-space:pre">	</span>         }
<span style="white-space:pre">		</span>matches_1.swap(inliers);
<span style="white-space:pre">		</span>std::cout << matches_1.size() << std::endl;

匹配點對根據RANSAC計算單應矩陣，再根據單應矩陣提純。