一、概述
- 行人检测过去流行采用的方法是DPM方法,其主要采用hog特征+SVM分类实现行人检测;
- 其中梯度方向直方图( Histogram of Oriented Gradients,HOG)的概念是
Dalal和Triggs在2005年提出,并将其用于行人检测,- 该方法在 MIT行人数据库上获得近乎 100% 的检测成功率;在包含视角、 光照和背景等变化的 INRIA 行人数据库上,也取得了大约 90% 的检测成功率。HOG是目前使用最为广泛的行人特征描述子。
参考博客:
二、实现代码
先上完整的代码,再进行解析说明
#include <iostream>
#include <fstream>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/objdetect/objdetect.hpp>
#include <opencv2/ml/ml.hpp>
#include <sys/time.h>
using namespace std;
using namespace cv;
using namespace cv::ml;
#define PosSamNO 2416 //original positive num
#define NegSamNO 6070 // original negative num
#define cropNegNum 1214 //number of to be croped picture
#define HardExampleNO 0 // hard negative num
#define AugPosSamNO 0 //Aug positive num
#define TRAIN true //if TRAIN is true, it will Train the data, if false it will not
#define CENTRAL_CROP true // it is nessary to set it to true,to fit the hog parameter
#define crop_negsample false //if true,it will crop the negtive sample at random
/********************************* 随机剪裁负样本 *******************************************/
void crop_negsample_random()
{
string imgName;
char saveName[200];
ifstream fileNeg("../img_dir/sample_neg.txt");
int num=0;
//如果文件存在,则先删除该文件
ofstream fout("../img_dir/sample_new_neg.txt",ios::trunc);
//读取负样本
for (int i = 0;i < cropNegNum && getline(fileNeg, imgName); i++)
{
imgName = "../normalized_images/train/neg/" + imgName; //加路径
Mat img = imread(imgName, IMREAD_UNCHANGED);
struct timeval tv;
if (img.empty())
{
cout << "can not load the image:" << imgName << endl;
continue;
}
if (img.cols >= 64 && img.rows >= 128)
{
num = 0;
//从每张图片中随机剪裁5张64*128的负样本
for (int j = 0;j < 5;j++)
{
gettimeofday(&tv,NULL);
srand(tv.tv_usec);//利用系统时间(微妙),设置随机数种子
int x = rand() % (img.cols - 64); //左上角x
int y = rand() % (img.rows - 128); //左上角y
cout << "x:" << x << "y:" << y <<endl;
Mat src = img(Rect(x, y, 64, 128));
sprintf(saveName, "../normalized_images/train/new_neg/neg%dCropped%d.png",i, num);
imwrite(saveName,src);
//保存裁剪得到的图片名称到txt文件,换行分隔
if(i<(cropNegNum-1)){
fout <<"neg" << i << "Cropped"<< num++ << ".png"<< endl;
}
else if(i==(cropNegNum-1) && j<4){
fout <<"neg" << i << "Cropped"<< num++ << ".png"<< endl;
}
else{
fout <<"neg" << i << "Cropped"<< num++ << ".png";
}
}
}
}
fout.close();
cout << "crop ok!" << endl;
}
int main()
{
if(crop_negsample){
crop_negsample_random(); //裁剪负样本
}
//winsize(64,128),blocksize(16,16),blockstep(8,8),cellsize(8,8),bins9
HOGDescriptor hog(Size(64,128),Size(16,16),Size(8,8),Size(8,8),9);
int DescriptorDim;
Ptr<SVM> svm = SVM::create();
if(TRAIN)
{
string ImgName;
ifstream finPos("../img_dir/sample_pos.txt");
// ifstream finNeg("../sample_neg.txt");
ifstream finNeg("../img_dir/sample_new_neg.txt");
if (!finPos || !finNeg)
{
cout << "Pos/Neg imglist reading failed..." << endl;
return 1;
}
Mat sampleFeatureMat;
Mat sampleLabelMat;
//loading original positive examples...
for(int num=0; num < PosSamNO && getline(finPos,ImgName); num++)
{
//cout <<"Now processing original positive image: " << ImgName << endl;
ImgName = "../normalized_images/train/pos/" + ImgName;
Mat src = imread(ImgName);
if(CENTRAL_CROP)
resize(src,src,Size(64,128));// src = src(Rect(16,16,64,128));
vector<float> descriptors;
hog.compute(src, descriptors, Size(8,8));//计算HOG描述子,检测窗口移动步长(8,8)
if( 0 == num )
{
DescriptorDim = descriptors.size();
sampleFeatureMat = Mat::zeros(PosSamNO +AugPosSamNO +NegSamNO +HardExampleNO, DescriptorDim, CV_32FC1);
sampleLabelMat = Mat::zeros(PosSamNO +AugPosSamNO +NegSamNO +HardExampleNO, 1, CV_32SC1);//sampleLabelMat的数据类型必须为有符号整数型
}
//将计算好的HOG描述子复制到样本特征矩阵sampleFeatureMat
for(int i=0; i<DescriptorDim; i++)
sampleFeatureMat.at<float>(num,i) = descriptors[i];
sampleLabelMat.at<int>(num,0) = 1;
}
finPos.close();
//loading original negative examples...
for(int num = 0; num < NegSamNO && getline(finNeg,ImgName); num++)
{
//cout<<"Now processing original negative image: "<<ImgName<<endl;
// ImgName = "../normalized_images/train/neg/" + ImgName;
ImgName = "../normalized_images/train/new_neg/" + ImgName;
Mat src = imread(ImgName);
vector<float> descriptors;
hog.compute(src,descriptors,Size(8,8));
for(int i=0; i<DescriptorDim; i++)
sampleFeatureMat.at<float>(num+PosSamNO,i) = descriptors[i];
sampleLabelMat.at<int>(num +PosSamNO +AugPosSamNO, 0) = -1;
}
finNeg.close();
svm ->setType(SVM::C_SVC);
svm ->setC(0.01);
svm ->setKernel(SVM::LINEAR);
// svm ->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER, 3000, 1e-6));
svm ->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER, 100, 1e-3));
cout<<"Starting training..."<<endl;
svm ->train(sampleFeatureMat, ROW_SAMPLE, sampleLabelMat);
cout<<"Finishing training..."<<endl;
svm ->save("../data/SVM_HOG.xml");
}
else {
svm = SVM::load( "../data/SVM_HOG.xml" );
}
cout << "loaded SVM_HOG.xml file" << endl;
Mat svecsmat = svm ->getSupportVectors();//svecsmat元素的数据类型为float
int svdim = svm ->getVarCount();//特征向量位数
int numofsv = svecsmat.rows;
// Mat alphamat = Mat::zeros(numofsv, svdim, CV_32F);//alphamat和svindex必须初始化,否则getDecisionFunction()函数会报错
Mat alphamat = Mat::zeros(numofsv, svdim, CV_32F);
Mat svindex = Mat::zeros(1, numofsv,CV_64F);
cout << "after initialize the value of alphamat is " << alphamat.size() << endl;
Mat Result;
double rho = svm ->getDecisionFunction(0, alphamat, svindex);
cout << "the value of rho is " << rho << endl;
alphamat.convertTo(alphamat, CV_32F);//将alphamat元素的数据类型重新转成CV_32F
cout << "the value of alphamat is " << alphamat << endl;
cout << "the size of alphamat is " << alphamat.size() << endl;
cout << "the size of svecsmat is " << svecsmat.size() << endl;
Result = -1 * alphamat * svecsmat;//float
cout << "the value of svdim is " << svdim << endl;
vector<float> vec;
for (int i = 0; i < svdim; ++i)
{
vec.push_back(Result.at<float>(0, i));
}
vec.push_back(rho);
cout << "going to write the HOGDetectorForOpenCV.txt file" << endl;
//saving HOGDetectorForOpenCV.txt
ofstream fout("HOGDetectorForOpenCV.txt");
for (int i = 0; i < vec.size(); ++i)
{
fout << vec[i] << endl;
}
/*********************************Testing**************************************************/
HOGDescriptor hog_test;
hog_test.setSVMDetector(vec);
// Mat src = imread("../person_and_bike_177b.png");
Mat src = imread("../3.jpg");
vector<Rect> found, found_filtered;
hog_test.detectMultiScale(src, found, 0, Size(8,8), Size(32,32), 1.05, 2);
cout<<"found.size : "<<found.size()<<endl;
//找出所有没有嵌套的矩形框r,并放入found_filtered中,如果有嵌套的话,则取外面最大的那个矩形框放入found_filtered中
for(int i=0; i < found.size(); i++)
{
Rect r = found[i];
int j=0;
for(; j < found.size(); j++)
if(j != i && (r & found[j]) == r)
break;
if( j == found.size())
found_filtered.push_back(r);
}
//画矩形框,因为hog检测出的矩形框比实际人体框要稍微大些,所以这里需要做一些调整
for(int i=0; i<found_filtered.size(); i++)
{
Rect r = found_filtered[i];
r.x += cvRound(r.width*0.1);
r.width = cvRound(r.width*0.8);
r.y += cvRound(r.height*0.07);
r.height = cvRound(r.height*0.8);
rectangle(src, r.tl(), r.br(), Scalar(0,255,0), 3);
}
imwrite("ImgProcessed.jpg",src);
namedWindow("src",0);
imshow("src",src);
waitKey(0);
return 0;
}
代码说明
- 程序在最开始的地方定义了负样本裁剪函数:crop_negsample_random;用于从少量的负样本中裁剪得到足够的负样本
- 行人数据集:INRIADATA,使用整理过的图片即:original_images这个文件夹中图片即可
- 当报错的时候,首先检查:正负样本的数量和对应的txt文件对应的行数是否一致
- 代码运行环境为:opencv3.0以上,在Ubuntu14.04+opencv3.2上运行测试,采用cmake进行编译进行了验证,目录组织如下图所示:
其中normalized_images为训练所需的图片,img_dir包含正负样本名称的txt文件,data包含一些测试的图片,完整代码可参考:https://download.csdn.net/download/yph001/10721973 - 其中在Ubuntu下将目录中的所有文件名输出到txt文件的方法为:
//例如将neg这个文件夹下面的所有文件名称,输入到与neg文件夹同级的sample_neg.txt目录
cd neg
ls -R * > ../sample_neg.txt
- 注意要将CENTRAL_CROP 设置为true,使得对正样本图片进行裁剪成64*128大小,使得这与hog参数中设置的(64,128)相吻合,要不很容易运行特别长时间,并最终报错