Opencv中HOG+SVM进行行人检测

一.基本函数介绍：

1.HOG特征提取：cv::HOGDescriptor类

CV_WRAP HOGDescriptor() :   
   winSize(64,128),                             // detect window    
   blockSize(16,16),                            // block 大小    
   blockStride(8,8),                            // overlap block的滑动步长    
   cellSize(8,8),                               // cell 大小     
   nbins(9),                                    // 直方图的bin个数    
   derivAperture(1),                            // 微分算子核    
   winSigma(-1),                                // 在window上进行高斯加权    
   histogramNormType(HOGDescriptor::L2Hys),     // 直方图归一化类型    
   L2HysThreshold(0.2),                         // L2 norm followed by clipping(limiting maximum values  of v to 0.                                                   2) and renormalising
   gammaCorrection(true),                       // Gamma校正，去除光照影响   
   nlevels(HOGDescriptor::DEFAULT_NLEVELS)      // 分层数  

2.一个window内HOG特征的维数计算：

(1) block的个数：

blockNum=[(winSize.x-blockStride.x)/blockStride.x+1]*[(winSize.y - blockStride.y)/blockStride.y+1]

(2) 一个block中特征维数：

featuresPerBlock = ( blockSize.x/cellSize.x)*( blockSize.y/cellSize.y)*nbins

(3) window内HOG特征维数：

blockNum*featuresPerBlock
则按上面默认参数计算，一个window内维数应该为3780维。

 

3.HOG特征提取源代码：

#include
#include
#include
#include "opencv2\opencv.hpp"
#define IMG_PATH "D:\\1.jpg"//图片路径
using namespace std;
int main(void) {
    cout << "hi!" << endl;
    cv::Mat mat = cv::imread(IMG_PATH);
     cout<<"width:"<<mat.cols<<" height:"<<mat.rows<<endl; //显示图片大小
    // 显示原图像
    cv::namedWindow("Image", CV_WINDOW_AUTOSIZE);
    cv::moveWindow("Image", 0, 0);
    cv::imshow("Image", mat);
    cv::HOGDescriptor hog;
    vector descriptors;
    hog.compute(mat, descriptors);
    cout<<"size of HOG:"<<descriptors.size()<<endl;  //显示图片提取的总的HOG特征维数
    cout << "finished." << endl;
    cv::waitKey(0);
    cout << "bye." << endl;
    return EXIT_SUCCESS;
}

二.HOG和SVM结合进行检测
1.HOG和SVM结合：
HOGDescriptor::setSVMDetector(const vector& detector)
其中 const vector& detector 在 opencv 中有已经训练好的行人检测分类器，直接用：
            HOGDescriptor::getDefaultPeopleDetector
            HOGDescriptor::getPeopleDetector48x96
            HOGDescriptor::getPeopleDetector64x128
 
2.对输入图像进行检测：
(1) 方法一：只在一个尺度上(即原图像上进行检测)
原型：
HOGDescriptor::detect(const Mat& img, vector& found_locations, double hit_threshold=0, Size win_stride=Size(), Size padding=Size())
参数解释：
  
img – Source image. CV_8UC1 and CV_8UC4 types are supported for now.
  
found_locations – Left-top corner points of detected objects boundaries.
  
hit_threshold – Threshold for the distance between features and SVM classifying plane. Usually it is 0 and should be specfied in the detector coefficients (as the last free coefficient). But if the free coefficient is omitted (which is allowed), you can specify it manually here.
  
win_stride – Window stride. It must be a multiple of block stride.
  
padding – Mock parameter to keep the CPU interface compatibility. It must be (0,0).
 
(2) 方法二：(在多个尺度上进行检测)
原型：
HOGDescriptor::detectMultiScale(const　Mat& img, vector& found_locations, doublehit_threshold=0, Size win_stride=Size(), Size padding=Size(), double scale0=1.05, int group_threshold=2)
参数解释：
  
img – Source image. See HOGDescriptor::detect() for type limitations.
  
found_locations – Detected objects boundaries.
  
hit_threshold – Threshold for the distance between features and SVM classifying plane. SeeHOGDescriptor::detect() for details.
  
win_stride – Window stride. It must be a multiple of block stride.
  
padding – Mock parameter to keep the CPU interface compatibility. It must be (0,0).
  
scale0 – Coefficient of the detection window increase.
  
group_threshold – Coefficient to regulate the similarity threshold. When detected, some objects can be covered by many rectangles. 0 means not to perform grouping. See groupRectangles()，说白了就是一个目标上覆盖有多个框的时候，通过设置该参数能去重(overlap)。
 
3.在图像上进行检测源代码
（参考了http://blog.youkuaiyun.com/icvpr/article/details/8454439的代码）

#include <iostream>   
#include <opencv2/opencv.hpp>   
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include "opencv2/objdetect/objdetect.hpp"
 int main()  
{  
        cv::Mat image = cv::imread("test.bmp");  
        if (image.empty())  
        {  
             std::cout<<"read image failed"<<std::endl;  
         }  
        // 定义HOG对象   
            cv::HOGDescriptor hog; // 采用默认参数   
        //  设置SVM分类器   采用已经训练好的行人检测分类器   
            hog.setSVMDetector(cv::HOGDescriptor::getDefaultPeopleDetector());   
        // 3. 在测试图像上检测行人区域   
            std::vector regions;  //存储返回的矩形框结果
        // 在对尺度上对Image进行检测
            hog.detectMultiScale(image, regions, 0, cv::Size(8,8), cv::Size(32,32), 1.05, 1);  
         // 显示检测后的结果  
         for (size_t i = 0; i < regions.size(); i++)  
         {  
                cv::rectangle(image, regions[i], cv::Scalar(0,0,255), 2);  
         }  
           cv::imshow("hog", image);  
           cv::waitKey(0);  
           return 0;  
 }  
 
 4.可能出错的地方：
 (1) hog.setSVMDetector(cv::HOGDescriptor::getDefaultPeopleDetector());   
改为：
static vector detector = HOGDescriptor::getDefaultPeopleDetector();
if (!detector.size())
{
       fprintf(stderr, "ERROR: getDefaultPeopleDetector returned NULL\n");
       return -1;     
}
hog.setSVMDetector(detector);
 
最后一个相关内容的比较好的英文网址：
http://docs.opencv.org/modules/gpu/doc/object_detection.html