一. RANSAC代码原理
:https://blog.csdn.net/robinhjwy/article/details/79174914
二. RANSAC的Opencv代码
这里只摘抄和RANSAC有关的部分,摘抄自:https://blog.csdn.net/Darlingqiang/article/details/79775542
- OpenCV中此功能通过调用findHomography函数调用,下面是个例程:
[cpp] view plain copy
#include <iostream>
#include "opencv2/opencv.hpp"
#include "opencv2/core/core.hpp"
#include "opencv2/features2d/features2d.hpp"
#include "opencv2/highgui/highgui.hpp"
using namespace cv;
using namespace std;
int main(int argc, char** argv)
{
Mat obj=imread("F:\\Picture\\obj.jpg"); //载入目标图像
Mat scene=imread("F:\\Picture\\scene.jpg"); //载入场景图像
if (obj.empty() || scene.empty() )
{
cout<<"Can't open the picture!\n";
return 0;
}
vector<KeyPoint> obj_keypoints,scene_keypoints;
Mat obj_descriptors,scene_descriptors;
ORB detector; //采用ORB算法提取特征点
detector.detect(obj,obj_keypoints);
detector.detect(scene,scene_keypoints);
detector.compute(obj,obj_keypoints,obj_descriptors);
detector.compute(scene,scene_keypoints,scene_descriptors);
BFMatcher matcher(NORM_HAMMING,true); //汉明距离做为相似度度量
vector<DMatch> matches;
matcher.match(obj_descriptors, scene_descriptors, matches);
Mat match_img;
drawMatches(obj,obj_keypoints,scene,scene_keypoints,matches,match_img);
imshow("滤除误匹配前",match_img);
//保存匹配对序号
vector<int> queryIdxs( matches.size() ), trainIdxs( matches.size() );
for( size_t i = 0; i < matches.size(); i++ )
{
queryIdxs[i] = matches[i].queryIdx;
trainIdxs[i] = matches[i].trainIdx;
}
Mat H12; //变换矩阵
vector<Point2f> points1; KeyPoint::convert(obj_keypoints, points1, queryIdxs);
vector<Point2f> points2; KeyPoint::convert(scene_keypoints, points2, trainIdxs);
int ransacReprojThreshold = 5; //拒绝阈值
H12 = findHomography( Mat(points1), Mat(points2), CV_RANSAC, ransacReprojThreshold );
vector<char> matchesMask( matches.size(), 0 );
Mat points1t;
perspectiveTransform(Mat(points1), points1t, H12);
for( size_t i1 = 0; i1 < points1.size(); i1++ ) //保存‘内点’
{
if( norm(points2[i1] - points1t.at<Point2f>((int)i1,0)) <= ransacReprojThreshold ) //给内点做标记
{
matchesMask[i1] = 1;
}
}
Mat match_img2; //滤除‘外点’后
drawMatches(obj,obj_keypoints,scene,scene_keypoints,matches,match_img2,Scalar(0,0,255),Scalar::all(-1),matchesMask);
//画出目标位置
std::vector<Point2f> obj_corners(4);
obj_corners[0] = cvPoint(0,0); obj_corners[1] = cvPoint( obj.cols, 0 );
obj_corners[2] = cvPoint( obj.cols, obj.rows ); obj_corners[3] = cvPoint( 0, obj.rows );
std::vector<Point2f> scene_corners(4);
perspectiveTransform( obj_corners, scene_corners, H12);
line( match_img2, scene_corners[0] + Point2f(static_cast<float>(obj.cols), 0),
scene_corners[1] + Point2f(static_cast<float>(obj.cols), 0),Scalar(0,0,255),2);
line( match_img2, scene_corners[1] + Point2f(static_cast<float>(obj.cols), 0),
scene_corners[2] + Point2f(static_cast<float>(obj.cols), 0),Scalar(0,0,255),2);
line( match_img2, scene_corners[2] + Point2f(static_cast<float>(obj.cols), 0),
scene_corners[3] + Point2f(static_cast<float>(obj.cols), 0),Scalar(0,0,255),2);
line( match_img2, scene_corners[3] + Point2f(static_cast<float>(obj.cols), 0),
scene_corners[0] + Point2f(static_cast<float>(obj.cols), 0),Scalar(0,0,255),2);
imshow("滤除误匹配后",match_img2);
waitKey(0);
return 0;
}
- RANSAC源码解析
(1)findHomography内部调用cvFindHomography函数
srcPoints:目标图像点集
dstPoints:场景图像点集
method: 最小中值法、RANSAC方法、最小二乘法
ransacReprojTheshold:投影误差阈值
mask:掩码
[cpp] view plain copy
cvFindHomography( const CvMat* objectPoints, const CvMat* imagePoints,
CvMat* __H, int method, double ransacReprojThreshold,
CvMat* mask )
{
const double confidence = 0.995; //置信度
const int maxIters = 2000; //迭代最大次数
const double defaultRANSACReprojThreshold = 3; //默认拒绝阈值
bool result = false;
Ptr<CvMat> m, M, tempMask;
double H[9];
CvMat matH = cvMat( 3, 3, CV_64FC1, H ); //变换矩阵
int count;
CV_Assert( CV_IS_MAT(imagePoints) && CV_IS_MAT(objectPoints) );
count = MAX(imagePoints->cols, imagePoints->rows);
CV_Assert( count >= 4 ); //至少有4个样本
if( ransacReprojThreshold <= 0 )
ransacReprojThreshold = defaultRANSACReprojThreshold;
m = cvCreateMat( 1, count, CV_64FC2 ); //转换为齐次坐标
cvConvertPointsHomogeneous( imagePoints, m );
M = cvCreateMat( 1, count, CV_64FC2 );//转换为齐次坐标
cvConvertPointsHomogeneous( objectPoints, M );
if( mask )
{
CV_Assert( CV_IS_MASK_ARR(mask) && CV_IS_MAT_CONT(mask->type) &&
(mask->rows == 1 || mask->cols == 1) &&
mask->rows*mask->cols == count );
}
if( mask || count > 4 )
tempMask = cvCreateMat( 1, count, CV_8U );
if( !tempMask.empty() )
cvSet( tempMask, cvScalarAll(1.) );
CvHomographyEstimator estimator(4);
if( count == 4 )
method = 0;
if( method == CV_LMEDS ) //最小中值法
result = estimator.runLMeDS( M, m, &matH, tempMask, confidence, maxIters );
else if( method == CV_RANSAC ) //采用RANSAC算法
result = estimator.runRANSAC( M, m, &matH, tempMask, ransacReprojThreshold, confidence, maxIters);//(2)解析
else
result = estimator.runKernel( M, m, &matH ) > 0; //最小二乘法
if( result && count > 4 )
{
icvCompressPoints( (CvPoint2D64f*)M->data.ptr, tempMask->data.ptr, 1, count ); //保存内点集
count = icvCompressPoints( (CvPoint2D64f*)m->data.ptr, tempMask->data.ptr, 1, count );
M->cols = m->cols = count;
if( method == CV_RANSAC ) //
estimator.runKernel( M, m, &matH ); //内点集上采用最小二乘法重新估算变换矩阵
estimator.refine( M, m, &matH, 10 );//
}
if( result )
cvConvert( &matH, __H ); //保存变换矩阵
if( mask && tempMask )
{
if( CV_ARE_SIZES_EQ(mask, tempMask) )
cvCopy( tempMask, mask );
else
cvTranspose( tempMask, mask );
}
return (int)result;
}
(2) runRANSAC
maxIters:最大迭代次数
m1,m2 :数据样本
model:变换矩阵
reprojThreshold:投影误差阈值
confidence:置信度 0.995
[cpp] view plain copy
bool CvModelEstimator2::runRANSAC( const CvMat* m1, const CvMat* m2, CvMat* model,
CvMat* mask0, double reprojThreshold,
double confidence, int maxIters )
{
bool result = false;
cv::Ptr<CvMat> mask = cvCloneMat(mask0);
cv::Ptr<CvMat> models, err, tmask;
cv::Ptr<CvMat> ms1, ms2;
int iter, niters = maxIters;
int count = m1->rows*m1->cols, maxGoodCount = 0;
CV_Assert( CV_ARE_SIZES_EQ(m1, m2) && CV_ARE_SIZES_EQ(m1, mask) );
if( count < modelPoints )
return false;
models = cvCreateMat( modelSize.height*maxBasicSolutions, modelSize.width, CV_64FC1 );
err = cvCreateMat( 1, count, CV_32FC1 );//保存每组点对应的投影误差
tmask = cvCreateMat( 1, count, CV_8UC1 );
if( count > modelPoints ) //多于4个点
{
ms1 = cvCreateMat( 1, modelPoints, m1->type );
ms2 = cvCreateMat( 1, modelPoints, m2->type );
}
else //等于4个点
{
niters = 1; //迭代一次
ms1 = cvCloneMat(m1); //保存每次随机找到的样本点
ms2 = cvCloneMat(m2);
}
for( iter = 0; iter < niters; iter++ ) //不断迭代
{
int i, goodCount, nmodels;
if( count > modelPoints )
{
//在(3)解析getSubset
bool found = getSubset( m1, m2, ms1, ms2, 300 ); //随机选择4组点,且三点之间不共线,(3)解析
if( !found )
{
if( iter == 0 )
return false;
break;
}
}
nmodels = runKernel( ms1, ms2, models ); //估算近似变换矩阵,返回1
if( nmodels <= 0 )
continue;
for( i = 0; i < nmodels; i++ )//执行一次
{
CvMat model_i;
cvGetRows( models, &model_i, i*modelSize.height, (i+1)*modelSize.height );//获取3×3矩阵元素
goodCount = findInliers( m1, m2, &model_i, err, tmask, reprojThreshold ); //找出内点,(4)解析
if( goodCount > MAX(maxGoodCount, modelPoints-1) ) //当前内点集元素个数大于最优内点集元素个数
{
std::swap(tmask, mask);
cvCopy( &model_i, model ); //更新最优模型
maxGoodCount = goodCount;
//confidence 为置信度默认0.995,modelPoints为最少所需样本数=4,niters迭代次数
niters = cvRANSACUpdateNumIters( confidence, //更新迭代次数,(5)解析
(double)(count - goodCount)/count, modelPoints, niters );
}
}
}
if( maxGoodCount > 0 )
{
if( mask != mask0 )
cvCopy( mask, mask0 );
result = true;
}
return result;
}
step one niters最初的值为2000,这就是初始时的RANSAC算法的循环次数,getSubset()函数是从一组对应的序列中随机的选出4组(因为要想计算出一个3X3的矩阵,至少需要4组对应的坐标),m1和m2是我们输入序列,ms1和ms2是随机选出的对应的4组匹配。
随机的选出4组匹配后,就应该根据这4个匹配计算相应的矩阵,所以函数runKernel()就是根据4组匹配计算矩阵,参数里的models就是得到的矩阵。
stept wo 这个矩阵只是一个假设,为了验证这个假设,需要用其他的点去计算,看看其他的点是内点还是外点。
findInliers()函数就是用来计算内点的。利用前面得到的矩阵,把所有的序列带入,计算得出哪些是内点,哪些是外点,函数的返回值为goodCount,就是此次计算的内点的个数。
step three 函数中还有一个值为maxGoodCount,每次循环的内点个数的最大值保存在这个值中,一个估计的矩阵如果有越多的内点,那么这个矩阵就越有可能是正确的。
step four 所以计算内点个数以后,紧接着判断一下goodCount和maxGoodCount的大小关系,如果goodCount>maxGoodCount,则把goodCount赋值给maxGoodCount。赋值之后的一行代码非常关键,我们单独拿出来说一下,代码如下:
niters = cvRANSACUpdateNumIters( confidence,
(double)(count - goodCount)/count, modelPoints, niters );
niters本来是迭代的次数,也就是循环的次数。但是通过这行代码我们发现,每次循环后,都会对niters这个值进行更新,也就是每次循环后都会改变循环的总次数。
cvRANSACUpdateNumIters()函数利用confidence(置信度)count(总匹配个数)goodCount(当前内点个数)niters(当前的总迭代次数)这几个参数,来动态的改变总迭代次数的大小。该函数的中心思想就是当内点占的比例较多时,那么很有可能已经找到了正确的估计,所以就适当的减少迭代次数来节省时间。这个迭代次数的减少是以指数形式减少的,所以节省的时间开销也是非常的可观。因此最初设计的2000的迭代次数,可能最终的迭代次数只有几十。同样的,如果你自己一开始把迭代次数设置成10000或者更大,进过几次迭代后,niters又会变得非常小了。所以初始时的niters设置的再大,其实对最终的运行时间也没什么影响。
所以,们现在应该清楚为什么输入数据增加,而算法运行时间不会增加了。openCV的RANSAC算法首先把迭代的次数设置为2000,然后再迭代的过程中,动态的改变总迭代次数,无论输入数据有多少,总的迭代次数不会增加,并且通过4个匹配计算出估计的矩阵这个时间是不变的,通过估计矩阵来计算内点,这方面的增加的时间开销基本上可以忽略。所以导致的最终结果就是,无论输入点有多少,运算时间基本不会有太大变化。
以上就是RANSAC算法的核心代码,其中用到的一些函数,下面一一给出。
(3)getSubset
ms1,ms2:保存随机找到的4个样本
maxAttempts:最大迭代次数,300
[cpp] view plain copy
bool CvModelEstimator2::getSubset( const CvMat* m1, const CvMat* m2,
CvMat* ms1, CvMat* ms2, int maxAttempts )
{
cv::AutoBuffer<int> _idx(modelPoints); //modelPoints 所需要最少的样本点个数
int* idx = _idx;
int i = 0, j, k, idx_i, iters = 0;
int type = CV_MAT_TYPE(m1->type), elemSize = CV_ELEM_SIZE(type);
const int *m1ptr = m1->data.i, *m2ptr = m2->data.i;
int *ms1ptr = ms1->data.i, *ms2ptr = ms2->data.i;
int count = m1->cols*m1->rows;
assert( CV_IS_MAT_CONT(m1->type & m2->type) && (elemSize % sizeof(int) == 0) );
elemSize /= sizeof(int); //每个数据占用字节数
for(; iters < maxAttempts; iters++)
{
for( i = 0; i < modelPoints && iters < maxAttempts; )
{
idx[i] = idx_i = cvRandInt(&rng) % count; // 随机选取1组点
for( j = 0; j < i; j++ ) // 检测是否重复选择
if( idx_i == idx[j] )
break;
if( j < i )
continue; //重新选择
for( k = 0; k < elemSize; k++ ) //拷贝点数据
{
ms1ptr[i*elemSize + k] = m1ptr[idx_i*elemSize + k];
ms2ptr[i*elemSize + k] = m2ptr[idx_i*elemSize + k];
}
if( checkPartialSubsets && (!checkSubset( ms1, i+1 ) || !checkSubset( ms2, i+1 )))//检测点之间是否共线
{
iters++; //若共线,重新选择一组
continue;
}
i++;
}
if( !checkPartialSubsets && i == modelPoints &&
(!checkSubset( ms1, i ) || !checkSubset( ms2, i )))
continue;
break;
}
return i == modelPoints && iters < maxAttempts; //返回找到的样本点个数
}
(4) findInliers & computerReprojError
[cpp] view plain copy
int CvModelEstimator2::findInliers( const CvMat* m1, const CvMat* m2,
const CvMat* model, CvMat* _err,
CvMat* _mask, double threshold )
{
int i, count = _err->rows*_err->cols, goodCount = 0;
const float* err = _err->data.fl;
uchar* mask = _mask->data.ptr;
computeReprojError( m1, m2, model, _err ); //计算每组点的投影误差
threshold *= threshold;
for( i = 0; i < count; i++ )
goodCount += mask[i] = err[i] <= threshold;//误差在限定范围内,加入‘内点集’
return goodCount;
}
//--------------------------------------
void CvHomographyEstimator::computeReprojError( const CvMat* m1, const CvMat* m2,const CvMat* model, CvMat* _err )
{
int i, count = m1->rows*m1->cols;
const CvPoint2D64f* M = (const CvPoint2D64f*)m1->data.ptr;
const CvPoint2D64f* m = (const CvPoint2D64f*)m2->data.ptr;
const double* H = model->data.db;
float* err = _err->data.fl;
for( i = 0; i < count; i++ ) //保存每组点的投影误差,对应上述变换公式
{
double ww = 1./(H[6]*M[i].x + H[7]*M[i].y + 1.);
double dx = (H[0]*M[i].x + H[1]*M[i].y + H[2])*ww - m[i].x;
double dy = (H[3]*M[i].x + H[4]*M[i].y + H[5])*ww - m[i].y;
err[i] = (float)(dx*dx + dy*dy);
}
}
(5)cvRANSACUpdateNumIters
对应上述k的计算公式
p:置信度
ep:外点比例
[cpp] view plain copy
cvRANSACUpdateNumIters( double p, double ep,
int model_points, int max_iters )
{
if( model_points <= 0 )
CV_Error( CV_StsOutOfRange, "the number of model points should be positive" );
p = MAX(p, 0.);
p = MIN(p, 1.);
ep = MAX(ep, 0.);
ep = MIN(ep, 1.);
// avoid inf's & nan's
double num = MAX(1. - p, DBL_MIN); //num=1-p,做分子
double denom = 1. - pow(1. - ep,model_points);//做分母
if( denom < DBL_MIN )
return 0;
num = log(num);
denom = log(denom);
return denom >= 0 || -num >= max_iters*(-denom) ?
max_iters : cvRound(num/denom);
}
参考https://blog.csdn.net/laobai1015/article/details/51683076
cv::findFundamentalMat
在处理立体图像对的时候经常会用到对极几何的知识,计算基础矩阵也是很常见的事。OpenCV实现了基本矩阵的算法。对于老版本的C代码,计算基本矩阵的RANSAC方法中有一个迭代次数不收敛的bug,可能导致每次计算的采样次数都达到最大限制的次数,其根源在于计算采样可信度的公式有误,新版本的代码还没有仔细看过,不敢确定是否已经修正了这个bug。但是这个bug并不会对最后的结果造成多大影响,只是会影响计算的效率——原本几次采样即可结束迭代,在这个bug的影响下可能要采样数百次。
新版本的计算基本矩阵的函数的使用也有一些问题,下面来看看cv::findFundamentalMat函数:
//! the algorithm for finding fundamental matrix
enum
{
FM_7POINT = CV_FM_7POINT, //!< 7-point algorithm
FM_8POINT = CV_FM_8POINT, //!< 8-point algorithm
FM_LMEDS = CV_FM_LMEDS, //!< least-median algorithm
FM_RANSAC = CV_FM_RANSAC //!< RANSAC algorithm
};
//! finds fundamental matrix from a set of corresponding 2D points
CV_EXPORTS Mat findFundamentalMat( const Mat& points1, const Mat& points2,
CV_OUT vector<uchar>& mask, int method=FM_RANSAC,
double param1=3., double param2=0.99 );
//! finds fundamental matrix from a set of corresponding 2D points
CV_EXPORTS_W Mat findFundamentalMat( const Mat& points1, const Mat& points2,
int method=FM_RANSAC,
double param1=3., double param2=0.99 );
上面是OpenCV计算基本矩阵的C++接口,其内部实现仍然是调用的C代码函数,仅仅是在上层进行了一次封装。网上有些文档中说const Mat& points1和points2这两个参数可以直接由vector类型作为传入参数,其实这是错误的。直接用vector传递,在编译时可能不会报错,但是运行时会直接崩溃。因为cv::Mat的构造函数并不会把Point2f转化为两个浮点数存于Mat的两个元素中,而是仍然以Point2f存于Mat的一个元素中,于是findFundamentalMat一读取这个Mat就会出错。
因此我们最好老老实实地去构建Mat points1和Mat points2。该矩阵的维度可以是2xN,也可以是Nx2的,其中N是特征点的数目。另一个要注意的地方就是该矩阵的类型不能是CV_64F,也就是说findFundamentalMat内部解析points1和points2时是按照float类型去解析的,设为CV_64F将导致读取数据失败,程序崩溃。最好设为CV_32F。下面是从匹配好的特征点计算F的代码示例:
// vector<KeyPoint> m_LeftKey;
// vector<KeyPoint> m_RightKey;
// vector<DMatch> m_Matches;
// 以上三个变量已经被计算出来,分别是提取的关键点及其匹配,下面直接计算F
// 分配空间
int ptCount = (int)m_Matches.size();
Mat p1(ptCount, 2, CV_32F);
Mat p2(ptCount, 2, CV_32F);
// 把Keypoint转换为Mat
Point2f pt;
for (int i=0; i<ptCount; i++)
{
pt = m_LeftKey[m_Matches[i].queryIdx].pt;
p1.at<float>(i, 0) = pt.x;
p1.at<float>(i, 1) = pt.y;
pt = m_RightKey[m_Matches[i].trainIdx].pt;
p2.at<float>(i, 0) = pt.x;
p2.at<float>(i, 1) = pt.y;
}
// 用RANSAC方法计算F
// Mat m_Fundamental;
// 上面这个变量是基本矩阵
// vector<uchar> m_RANSACStatus;
// 上面这个变量已经定义过,用于存储RANSAC后每个点的状态
m_Fundamental = findFundamentalMat(p1, p2, m_RANSACStatus, FM_RANSAC);
// 计算野点个数
int OutlinerCount = 0;
for (int i=0; i<ptCount; i++)
{
if (m_RANSACStatus[i] == 0) // 状态为0表示野点
{
OutlinerCount++;
}
}
// 计算内点
// vector<Point2f> m_LeftInlier;
// vector<Point2f> m_RightInlier;
// vector<DMatch> m_InlierMatches;
// 上面三个变量用于保存内点和匹配关系
int InlinerCount = ptCount - OutlinerCount;
m_InlierMatches.resize(InlinerCount);
m_LeftInlier.resize(InlinerCount);
m_RightInlier.resize(InlinerCount);
InlinerCount = 0;
for (int i=0; i<ptCount; i++)
{
if (m_RANSACStatus[i] != 0)
{
m_LeftInlier[InlinerCount].x = p1.at<float>(i, 0);
m_LeftInlier[InlinerCount].y = p1.at<float>(i, 1);
m_RightInlier[InlinerCount].x = p2.at<float>(i, 0);
m_RightInlier[InlinerCount].y = p2.at<float>(i, 1);
m_InlierMatches[InlinerCount].queryIdx = InlinerCount;
m_InlierMatches[InlinerCount].trainIdx = InlinerCount;
InlinerCount++;
}
}
// 把内点转换为drawMatches可以使用的格式
vector<KeyPoint> key1(InlinerCount);
vector<KeyPoint> key2(InlinerCount);
KeyPoint::convert(m_LeftInlier, key1);
KeyPoint::convert(m_RightInlier, key2);
// 显示计算F过后的内点匹配
// Mat m_matLeftImage;
// Mat m_matRightImage;
// 以上两个变量保存的是左右两幅图像
Mat OutImage;
drawMatches(m_matLeftImage, key1, m_matRightImage, key2, m_InlierMatches, OutImage);
cvNamedWindow( "Match features", 1);
cvShowImage("Match features", &(IplImage(OutImage)));
cvWaitKey( 0 );
cvDestroyWindow( "Match features" );
————————————————
版权声明:本文为CSDN博主「sunshine_guoqiang」的原创文章,遵循CC 4.0 BY-SA版权协议,转载请附上原文出处链接及本声明。
原文链接:https://blog.csdn.net/Darlingqiang/article/details/79775542