本文介绍了一种改进的OpenCV MeanShift算法在行人检测中的应用，通过局部融合和重叠矩形框的去除提高了检测精度。文章详细解释了如何通过计算矩形框的得分并基于得分进行融合来提高检测效果。
OPENCV 中的代码改进。当然要依据自己的实际情况来，OPENCV 中行人检測有两种矩形框的融合算法。这里仅仅对meanshift 方法做改进
假设有更好的方法。希望能够跟我讲下。
对于去除重合部分。我也写了改进，看懂了能够加到自己程序中。
为什么要做局部MeanShift?
图1.全局MeanShift
如图所看到的：两幅图像距离较近且有多个矩形框。全局MeanShift融合后可能会造成这样的结果
而假设用局部融合就能避免这样的情况。
/*----------------------------------分数判定：begin--------------------------------------------------*/
//////////////将感兴趣区域归一化后计算HOG特征--begin////////////////////
cvSetImageROI(imageOrigin,r);
IplImage *test=cvCreateImage(cvGetSize(imageOrigin),8,3);
cvCopyImage(imageOrigin,test);
cvResetImageROI(imageOrigin);
IplImage* testTempImg=cvCreateImage(cvSize(40,40),8,3);
cvZero(testTempImg);
cvResize(test,testTempImg);
hog->compute(testTempImg, descriptors,Size(1,1), Size(0,0)); //Hog特征计算
Mat another(descriptors);
descriptors.clear();
////////////将感兴趣区域归一化后计算HOG特征--end///////////////
double ret = 0;
ret = ssvm.get_calc_liner(vec_count,var_count,a,another,results,alpha,0)-rho; //计算SVM分数
////////////////////////////
cvReleaseImage(&test);
cvReleaseImage(&testTempImg);
////////////////////////////
if(ret <0)
{
continue; //去掉SVM 分值小于0的图像
}
rc_.push_back(rcc);
weights.push_back(ret);
double rate = max(rcc.width*1.0/40,rcc.height*1.0/40);
//选取长宽最大值，作为尺度，考虑还不够周到，这里改变。Meanshift 所有都要改
//这里还能够改进
foundScales.push_back(rate);
/*----------------------------------分数判定：end--------------------------------------------------*/
//////////上述应该是一个循环，增加多个点//////////////
/////////////////////////融合过程-begin////////////////////////////
//vector<Rect> rc_;//矩形框
// vector<double> weights;//权重，score
// vector<double> foundScales;//放缩尺度
groupRectangles_meanshift1(rc_, weights, foundScales, 0.3, Size(40,40)); //框出来的矩形框进行融合
//groupRectangles_meanshift1该函数在最后定义//
///////////////////////融合过程-end///////////////////////////
//////////////////重合去重第一步：计算融合后的分值//////////////////
for( i = 0; i < rc_.size(); i++ )
{
//增加矩形框
Rect r = rc_[i];
found_filtered.push_back(r);
}
vector<float> found_score(found_filtered.size()); //矩形框的分数
for( i = 0; i < found_filtered.size(); i++ )
{
Rect r = found_filtered[i];
//////////////////////////////////
cvSetImageROI(imageOrigin,r);
IplImage *test=cvCreateImage(cvGetSize(imageOrigin),8,3);
cvCopyImage(imageOrigin,test);
cvResetImageROI(imageOrigin);
IplImage* testTempImg=cvCreateImage(cvSize(40,40),8,3);
cvZero(testTempImg);
cvResize(test,testTempImg);
hog->compute(testTempImg, descriptors,Size(1,1), Size(0,0)); //Hog特征计算
Mat another(descriptors);
descriptors.clear();
double ret = 0;
ret= ssvm.get_calc_liner(vec_count,var_count,a,another,results,alpha,0)-rho;
cvReleaseImage(&test);
cvReleaseImage(&testTempImg);
found_score[i]=ret;
////////////////////////////
}
////////重合去重第二步：矩形框内的部分。取分值最大的/////
//////////////////////found_filtered为融合后的矩形框//////////////////////////
for( i = 0; i < found_filtered.size(); i++ )
{
//进行了重合去除，选取附近分值最大矩形框
Rect r = found_filtered[i];
for( j = 0; j < rc_.size(); j++ )
if( j != i && (r & found_filtered[j]).area() == r.area())
{//这里是将重叠的部分，分值小的矩形框的权重设为-1,为了取最大值
if(found_score[i]>found_score[j])
{
found_score[j]=-1;
break;
}
else
{
found_score[i]=-1;
break;
}
}
}
////////重合去重第三步：重叠部分。取分值最大的/////
//////////////////////////////////////////////////////////////////
for( i = 0; i < found_filtered.size(); i++ )
{
//进行了重合去除
Rect r = found_filtered[i];
for( j = 0; j < rc_.size(); j++ )
//判定重合是否大于相较面积的70%（这个比例有待測试）
//判定都为上一步过滤后的结果，可能存在部分相较。但不包括的情况
if( j != i && (r & found_filtered[j]).area() >= r.area()*0.7 &&found_score[j]!=-1&&found_score[i]!=-1)
{//这里是将重叠的部分，分值小的矩形框的权重设为-1
if(found_score[i]>found_score[j])
{
found_score[j]=-1;
break;
}
else
{
found_score[i]=-1;
break;
}
}
}
for(int i=0;i<found_filtered.size();i++)
{
if (found_score[i]==-1)//将分值等于-1的过滤掉
{
continue;
}
Rect r = found_filtered[i];
r.x -= cvRound(r.width*0.05);
r.width = cvRound(r.width*1.05);
r.y -= cvRound(r.height*0.05);
r.height = cvRound(r.height*1.05);
rectangle(img_dst, r.tl(), r.br(), cv::Scalar(0,255,0), 2); //在图像上画矩形框
}
/*-----------------------MeanShift做局部的（源程序是对全局）------------------------*/
//meanshift 融合
class MeanshiftGrouping
{
public:
// MeanshiftGrouping msGrouping(smothing, hits,rectList, hitWeights, 1e-5, 100);//得到
///////////////////////////////////////////////////////////////////////
// msGrouping.getModes(resultHits, resultWeights, 1);
//////////////////////////////////////////////////////////////////////
MeanshiftGrouping(const Point3d& densKer, const vector<Point3d>& posV,const vector<Rect>&list,
const vector<double>& wV, double eps, int maxIter = 20)
{
densityKernel = densKer;
weightsV = wV;
positionsV = posV;
positionsCount = (int)posV.size();
meanshiftV.resize(positionsCount);
distanceV.resize(positionsCount);
iterMax = maxIter;
modeEps = eps;
for (unsigned i = 0; i<positionsV.size(); i++)
{
meanshiftV[i] = getNewValue(positionsV[i],list[i],list);//positionV 仅仅有中点坐标没有长宽。distanceV[i] = moveToMode(meanshiftV[i],list[i],list);//做最大为iterMax次循环//均值漂移后的值 meanshiftV[i] -= positionsV[i];//这一步后面没用到 } } void getModes(vector<Point3d>& modesV, vector<double>& resWeightsV, const double eps) { for (size_t i=0; i <distanceV.size(); i++) { bool is_found = false; for(size_t j=0; j<modesV.size(); j++) { if ( getDistance(distanceV[i], modesV[j]) < eps)//欧式距离小于阈值 { is_found=true; break; } } if (!is_found) { modesV.push_back(distanceV[i]);//增加距离较大的点，也就是说两个点距离较大。不是同一个矩形框 } } resWeightsV.resize(modesV.size()); for (size_t i=0; i<modesV.size(); i++) { resWeightsV[i] = getResultWeight(modesV[i]);//得到点的权值 } } protected: vector<Point3d> positionsV; vector<double> weightsV; Point3d densityKernel; int positionsCount; vector<Point3d> meanshiftV; vector<Point3d> distanceV; int iterMax; double modeEps; Point3d getNewValue(const Point3d& inPt ,const Rect inR ,const vector<Rect>list) const {//inPt 输入三维坐标 、inR 为输入的矩形 list为所有矩形 Point3d resPoint(.0); Point3d ratPoint(.0); int value=20;//仅仅考虑矩形框四个角差值小于20的点。这个能够自己设定 for (size_t i=0; i<positionsV.size(); i++) { Point3d aPt= positionsV[i]; // double rate = exp(aPt.z); if(inR.x -list[i].x>value||inR.y -list[i].y>value||inR.width -list[i].width>value||inR.height -list[i].height>value) continue;//局部推断，假设不是同一个矩形附近，将排除附近矩形的影响 Point3d bPt = inPt; Point3d sPt = densityKernel;//核 //////////////////////////////////////// sPt.x *= exp(aPt.z);//Z为尺度 sPt.y *= exp(aPt.z); aPt.x /= sPt.x; aPt.y /= sPt.y; aPt.z /= sPt.z; bPt.x /= sPt.x; bPt.y /= sPt.y; bPt.z /= sPt.z; ///映射到相应尺度的图片的坐标/////////sPt为scale//反归一化 //////////////////////////////////////////// double w = (weightsV[i])*std::exp(-((aPt-bPt).dot(aPt-bPt))/2)/std::sqrt(sPt.dot(Point3d(1,1,1))); //又一次计算的权重，原权重weightsV[i]为线性SVM的得分 resPoint += w*aPt;//依据中心点的距离加相应的权值,距离越近,权值越大 ratPoint.x += w/sPt.x;//这边除以权重值，使得放缩后的图像权重变小 ratPoint.y += w/sPt.y; ratPoint.z += w/sPt.z; } resPoint.x /= ratPoint.x; resPoint.y /= ratPoint.y; resPoint.z /= ratPoint.z; return resPoint;//返回被周围点影响后的值 } double getResultWeight(const Point3d& inPt) const { double sumW=0;//终于返回的值 int num=0; size_t aa=positionsV.size();//位置点的个数 int len = int(aa);//位置点的个数 for (size_t i=0; i<aa; i++) { Point3d aPt = positionsV[i]; Point3d sPt = densityKernel; sPt.x *= exp(aPt.z); sPt.y *= exp(aPt.z); aPt -= inPt;//讲个坐标之差 aPt.x /= sPt.x; aPt.y /= sPt.y; aPt.z /= sPt.z; /*-----------------这块还能够优化，策略的考虑，权重的选取--begin-----------------*/ if(aa>10)//考虑假设aa的数量过小时 { double mm = aPt.dot(aPt); if(aPt.dot(aPt)<=0.5)//两个点的欧式距离 { if(weightsV[i]>0.6)//weightsV[i] 为svm的权值 { sumW+=0.35; } else if(weightsV[i]>0.3) { sumW+=0.3; } continue; } //////////////////////////原始/////////// sumW+=(weightsV[i])*std::exp(-(aPt.dot(aPt))/2)/std::sqrt(sPt.dot(Point3d(1,1,1))); ///////////////////////////////////////// } else { //sumW+=weightsV[i]; sumW+=(weightsV[i])*std::exp(-(aPt.dot(aPt))/2)*2.8; } /*---------待优化部分-----------------*/ /*------这块还能够优化。策略的考虑，权重的选取--end------*/ return sumW;//计算最后的权值 } Point3d moveToMode(Point3d aPt ,Rect inR, const vector<Rect>list) const {//均值漂移后的位置 Point3d bPt; for (int i = 0; i<iterMax; i++) { bPt = aPt; aPt = getNewValue(bPt,inR,list); if ( getDistance(aPt, bPt) <= modeEps )//小于阈值时返回，说明达到稳定状态 { break; } } return aPt;//返回稳定状态的值 } double getDistance(Point3d p1, Point3d p2) const {//计算欧式距离 Point3d ns = densityKernel; ns.x *= exp(p2.z); ns.y *= exp(p2.z); p2 -= p1; p2.x /= ns.x; p2.y /= ns.y; p2.z /= ns.z; return p2.dot(p2); } }; void groupRectangles_meanshift2(vector<Rect>& rectList, double detectThreshold, vector<double>* foundWeights, vector<double>& scales, Size winDetSize) //被groupRectangles_meanshift1调用 { int detectionCount = (int)rectList.size();//矩形的个数 vector<Point3d> hits(detectionCount), resultHits; vector<double> hitWeights(detectionCount), resultWeights; Point2d hitCenter; for (int i=0; i < detectionCount; i++) { hitWeights[i] = (*foundWeights)[i]; hitCenter = (rectList[i].tl() + rectList[i].br())*(0.5); //center of rectangles hits[i] = Point3d(hitCenter.x, hitCenter.y, std::log(scales[i]));//中心坐标x、y、log(scale) } if (foundWeights) foundWeights->clear(); double logZ = std::log(1.3); Point3d smothing(8, 16, logZ); MeanshiftGrouping msGrouping(smothing, hits,rectList, hitWeights, 1e-5, 100);//得到 /////////////////////////////////////////////////////////////////////// msGrouping.getModes(resultHits, resultWeights, 1); ////////////////////////////////////////////////////////////////////// rectList.clear(); for (unsigned i=0; i < resultHits.size(); ++i) { double scale = exp(resultHits[i].z);//还原尺度 hitCenter.x = resultHits[i].x;//中心坐标 hitCenter.y = resultHits[i].y; Size s( int(winDetSize.width*scale), int(winDetSize.height*scale) );//还原窗的大小 Rect resultRect( int(hitCenter.x-s.width/2), int(hitCenter.y-s.height/2), int(s.width), int(s.height) );//还原窗的位置 if (resultWeights[i] > detectThreshold)//detectThreshold 大于阈值的权重值输出 {//返回矩形框 和 权值 rectList.push_back(resultRect); foundWeights->push_back(resultWeights[i]); } } } void groupRectangles_meanshift1(vector<Rect>& rectList, vector<double>& foundWeights, vector<double>& foundScales, double detectThreshold, Size winDetSize) { groupRectangles_meanshift2(rectList, detectThreshold, &foundWeights, foundScales, winDetSize); //rectList矩形列表, detectThreshold阈值, foundWeights得分, foundScales尺度, winDetSize窗体大小 }
扫一扫
1899
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
