OpenCV HOG PeopleDetector：从训练到检测

【原文：http://blog.youtueye.com/work/opencv-hog-peopledetector-trainning.html】

opencv2.0 以上版本提供了行人检测的方法，以opencv2.2为例，该示例为opencv安装目录下的：/samples/cpp/peopledetect.cpp，测试效果如下：

我们先看看示例代码的内容：

#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/objdetect/objdetect.hpp"
#include "opencv2/highgui/highgui.hpp"

#include <stdio.h>
#include <string.h>
#include <ctype.h>

using namespace cv;
using namespace std;

void help()
{
    printf (
            "\nDemonstrate the use of the HoG descriptor using\n"
            "  HOGDescriptor::hog.setSVMDetector(HOGDescriptor::getDefaultPeopleDetector());\n"
            "Usage:\n"
            "./peopledetect (<image_filename> | <image_list>.txt)\n\n" );
}

int main( int argc, char ** argv)
{
    Mat img;
    FILE * f = 0;
    char _filename[1024];

    if ( argc == 1 )
    {
        printf ( "Usage: peopledetect (<image_filename> | <image_list>.txt)\n" );
        return 0;
    }
    img = imread(argv[1]);

    if ( img.data )
    {
        strcpy (_filename, argv[1]);
    }
    else
    {
        f = fopen (argv[1], "rt" );
        if (!f)
        {
            fprintf ( stderr, "ERROR: the specified file could not be loaded\n" );
            return -1;
        }
    }

    HOGDescriptor hog;
    hog.setSVMDetector(HOGDescriptor::getDefaultPeopleDetector());
    namedWindow( "people detector" , 1);

    for (;;)
    {
        char * filename = _filename;
        if (f)
        {
            if (! fgets (filename, ( int ) sizeof (_filename)-2, f))
                break ;
            //while(*filename && isspace(*filename))
            //  ++filename;
            if (filename[0] == '#' )
                continue ;
            int l = strlen (filename);
            while (l > 0 && isspace (filename[l-1]))
                --l;
            filename[l] = '\0' ;
            img = imread(filename);
        }
        printf ( "%s:\n" , filename);
        if (!img.data)
            continue ;

        fflush (stdout);
        vector<Rect> found, found_filtered;
        double t = ( double )getTickCount();
        // run the detector with default parameters. to get a higher hit-rate
        // (and more false alarms, respectively), decrease the hitThreshold and
        // groupThreshold (set groupThreshold to 0 to turn off the grouping completely).
        hog.detectMultiScale(img, found, 0, Size(8,8), Size(32,32), 1.05, 2);
        t = ( double )getTickCount() - t;
        printf ( "tdetection time = %gms\n" , t*1000./cv::getTickFrequency());
        size_t i, j;
        for ( i = 0; i < found.size(); i++ )
        {
            Rect r = found[i];
            for ( j = 0; j < found.size(); j++ )
                if ( j != i && (r & found[j]) == r)
                    break ;
            if ( j == found.size() )
                found_filtered.push_back(r);
        }
        for ( i = 0; i < found_filtered.size(); i++ )
        {
            Rect r = found_filtered[i];
            // the HOG detector returns slightly larger rectangles than the real objects.
            // so we slightly shrink the rectangles to get a nicer output.
            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(img, r.tl(), r.br(), cv::Scalar(0,255,0), 3);
        }
        imshow( "people detector" , img);
        int c = waitKey(0) & 255;
        if ( c == 'q' || c == 'Q' || !f)
            break ;
    }
    if (f)
        fclose (f);
    return 0;
}

你会发现代码中没有类似人脸检测时的分类器（模型）的载入过程，而是用

hog.setSVMDetector(HOGDescriptor::getDefaultPeopleDetector());

载入默认模型。找到hog的源码，在/modules/objdetect/hog.cpp，从第907行开始的代码如下：

vector< float > HOGDescriptor::getDefaultPeopleDetector()
{
    static const float detector[] = {
       0.05359386f, -0.14721455f, -0.05532170f, 0.05077307f,
       0.11547081f, -0.04268804f, 0.04635834f, -0.05468199f, 0.08232084f,
       0.10424068f, -0.02294518f, 0.01108519f, 0.01378693f, 0.11193510f,
       0.01268418f, 0.08528346f, -0.06309239f, 0.13054633f, 0.08100729f,
       -0.05209739f, -0.04315529f, 0.09341384f, 0.11035026f, -0.07596218f,
       -0.05517511f, -0.04465296f, 0.02947334f, 0.04555536f,
       -3.55954492e-003f, 0.07818956f, 0.07730991f, 0.07890715f, 0.06222893f,
       0.09001380f, -0.03574381f, 0.03414327f, 0.05677258f, -0.04773581f,
       0.03746637f, -0.03521175f, 0.06955440f, -0.03849038f, 0.01052293f,
       0.01736112f, 0.10867710f, 0.08748853f, 3.29739624e-003f, 0.10907028f,
       0.07913758f, 0.10393070f, 0.02091867f, 0.11594022f, 0.13182420f,
       0.09879354f, 0.05362710f, -0.06745391f, -7.01260753e-003f,....

显然，行人检测所用到的模型数据被做为常量写在源代码中，但这些模型数据数据是如何的来的呢？如果我要用我自己的样本tainning ，然后用的到的model去detect该如何操作呢？别急，我们从opencv行人检测所用到的算法说起。

opencv行人检测所用到的算法源自Navneet Dalal和Bill Triggs在2005's CVPR上的文章Histograms of Oriented Gradients for Human Detection.这里是英文和中文的介绍。作者所在的研究机构(INRIA：French National Institute for Research in Computer Science and Control,法国国家计算机技术和控制研究所)发布了这套算法的源码：INRIA Object Detection and Localization Toolkit.

那么，opencv中的PeopleDetector是否就原封不动地使用了这套算法呢，为了求证这一问题，笔者首先比较了两者的模型数据：下载OLTbinaries.zip，找到/HOG/model_4BiSVMLight.alt文件（二进制形式的数据），用代码将其中数据保存为文本格式（这个要根据上面的源码learcode.zip），其内容如下：

0.05359386 -0.14721455 -0.05532170 0.05077307 0.11547081 -0.04268804 0.04635834 -0.05468199 0.08232084 0.10424068
-0.02294518 0.01108519 0.01378693 0.11193510 0.01268418 0.08528346 -0.06309239 0.13054632 0.08100729 -0.05209739
-0.04315529 0.09341384 0.11035026 -0.07596217 -0.05517511 -0.04465296 0.02947334 0.04555536 -0.00355954 0.07818956
0.07730991 0.07890716 0.06222893 0.09001381 -0.03574381 0.03414327 0.05677258 -0.04773581 0.03746637 -0.03521175
0.06955440 -0.03849038 0.01052293 0.01736112 0.10867710 0.08748853 0.00329740 0.10907028 0.07913758 0.10393069
0.02091867 0.11594022 0.13182420 0.09879354 0.05362710 -0.06745391 -0.00701261 0.00524702 0.03236255 0.01407916
0.02207983 0.02537322 0.04547948 0.07200756 0.03129894 -0.06274468 0.02107014 0.06035208 0.08636236 0.00453164
0.02193363 0.02309801 0.05568166 -0.02645093 0.04448695 0.02837519 0.08975695 0.04461517 0.08975355 0.07514391
0.02306982 0.10410084 0.06368385 0.05943465 0.00458421 0.05220337 0.06675851 0.08358569 0.06712102 0.06559004
-0.03930481 -0.00915937 -0.05897915 0.02816453 0.05032348 0.06780671 0.03377650 -0.00060942 -0.01795146 -0.03083684
-0.01302475 -0.02972313 0.00788707 -0.03525961 -0.00250398 0.05245084 0.11791293 -0.02167498 0.05299332 0.06640524
0.05190265 -0.00827317 0.03033127 0.05842173 -0.00401050 -0.00625106 0.05862958 -0.02465461 0.05546781 -0.08228195 ....

数值居然同HOGDescriptor::getDefaultPeopleDetector()中的是一样！那么，你就可以用上面的OLT去trainning你的样本，然后把得到的model数据替换getDefaultPeopleDetector()中的数据，你便可以去进行你要的目标检测啦！

为了验证这一想法的正确性和可行性，笔者做了些实验，在Ubuntu10.4 g++4.4.5环境中，步骤如下：

1. 下载release版的程序：OLTbinaries.zip
2. 下载样本：INRIAPerson.tar
3. 在目录OLTbinaries/下建立两个文件夹test, train. 将INRIAperson/Test/neg拷贝到test/下，INRIAperson/Train/neg拷贝到train/下;将INRIAperson/test_64x128_H96拷贝到test/下重命名为pos，将INRIAperson/train_64x128_H96拷贝到train/下重命名为pos;
4. 将test/neg , test/pos各自文件夹中的所有图片文件名分别输出到neg.list, pos.list, 并放置在test/下; 同样地操作在train/。

   <pre class="brush: shell; gutter: true">amadeuzou@ubuntu:~/Projects/opencv/OLTbinaries/test$ ls ./neg &gt; neg.list
   amadeuzou@ubuntu:~/Projects/opencv/OLTbinaries/test$ ls ./pos &gt; pos.list</pre>

5. 到这里，样本数据便准备好了，那么，只要修改OLTbinaries/runall.sh相关参数然后这些此脚本，一小时左右的时间，便会在OLTbinaries/HOG/下产生一个model_4BiSVMLight.alt文件，你的模型数据便保存在这里面。到这里，你便成功trainning了一个model。

注意事项：

• runall.sh中第5行，按你的正负样本数目修改：

  HardOption= " --poscases 2416 --negcases 1218 "

• runall.sh中第21行，按你的样本文件夹所在(InDir)及输出文件所在(OutDir)修改：

  OutDir=. /HOG
  InDir=./
  OutFile=$OutDir /record
  CMDLINE=$OutDir /record

• trainning过程中会产生两个G左右的临时文件在OutDir（=./HOG）中，所以要确保硬盘空间足够，结束时删除，只留model_4BiSVMLight.alt。
• 整个trainning过程分4步，有4条屏幕输出信息，最可能出现的错误是样本文件路径不对，可在pos.list neg.list 中用图像文件的绝对路径。

  First iteration complete
  Hard examples created
  Doing second learning
  Second iteration complete

• 如果你用的是自己的样本，注意修改其他参数（待究），如正样本的大小：

  WIDTH=64; export WIDTH
  HEIGHT=128; export HEIGHT

有了模型，怎么去做目标检测呢？你可以做以下的试验：

1. 使用bin在OLTbinaries/bin/中classify_rhog： classify_rhog [待检测图片] [目标位置数据结果保存的文本文件] [模型文件] -i [位置画在图像文件]

   amadeuzou@ubuntu:~/Projects/opencv/OLTbinaries/bin$ ./classify_rhog  person-1.jpg  result.txt model_4BiSVMLight.alt -i result.jpg

   结果：

2. 使用lib在OLTbinaries/lib/中：

   /*
    * =============================================
    *
    *       Filename:  lib-detector.cpp
    *
    *    Description:  Code to detect object
    *
    *       Compiler:  gcc
    *
    *         Author:  Amadeu zou
    *         URL:  www.youtueye.com
    *
    * =============================================
    */
   #include <cv.h>
   #include <highgui.h>
   
   #include <string>
   #include <iostream>
   #include <algorithm>
   #include <iterator>
   
   #include <X11/Xlib.h>
   #include <Imlib2.h>
   
   #include <lear/interface/windetect.h>// change this path as appropriate.
   using namespace std;
   
   std::list<DetectedRegion> detector( char * imgf, WinDetectClassify windetect, LinearClassify* classifier)
   {
       std::list<DetectedRegion> detections;
       // read image
       Imlib_Image image = imlib_load_image(imgf);
       // if the load was successful
       if (image)
       { // set the image we loaded as the current context image to work on
           imlib_context_set_image(image);
       } else {
           //std::cerr << "Unable to read image: " << argv[1] << std::endl;
           return detections;
       }
       int width  = imlib_image_get_width(),
           height = imlib_image_get_height();
   
       typedef unsigned char uchar;
       DATA32* data = imlib_image_get_data_for_reading_only();
       uchar* imagedata = new uchar[3*width*height* sizeof (uchar)];
       for ( int j= 0; j< height; ++j)
       for ( int i= 0; i< width; ++i) {
           uchar* pixel = imagedata+(i+j*width)*3;
           int argb = data[i+j*width];
           pixel[0] = static_cast <uchar>((argb & 0x00FF0000)>>16);
           pixel[1] = static_cast <uchar>((argb & 0x0000FF00)>> 8);
           pixel[2] = static_cast <uchar>((argb & 0x000000FF)    );
       }
       imlib_free_image();
   
       // now get detections
       windetect.test(*classifier, detections, imagedata, width, height);
   
       delete [] imagedata;
       return detections;
   }
   
   int main( int argc, char ** argv) {
   
       if (argc != 4) {
           std::cout << "Error" << std::endl;
           return 0;
       }
   
       char modelpath[256];
       strcpy (modelpath,argv[2]);
       string model_file(modelpath) ;
       // initialize the person detector. All default parameters are set for person detector.
       WinDetectClassify windetect; // use default person detector.
       RHOGDenseParam desc;
       LinearClassify* classifier = NULL; // initialize it to 64x128 person detector.
       classifier = new LinearClassify(model_file, 0);
       windetect.init(&desc); // initialize the descriptor computation
   
       std::list<DetectedRegion> detections;
       detections = detector(argv[1], windetect, classifier);
   
       //draw region in image
       IplImage* img = cvLoadImage(argv[1],1);
       for (list<DetectedRegion>::iterator itor=detections.begin();itor!=detections.end();++itor)
       {
            cvRectangle(img,cvPoint(itor->x,itor->y),cvPoint(itor->x+itor->width,itor->y+itor->height),cvScalar(0,0,255),2);
   
       }
       cvSaveImage(argv[3],img);
       cvReleaseImage(&img);
       //print detections
       std::copy(detections.begin(), detections.end(), std::ostream_iterator<DetectedRegion>(std::cout, "\n" ));
   
       return 0;
   }

   编译：

   g++ `pkg-config --cflags --libs opencv` -O3 -o lib-detector lib-detector.cpp   -I.  -I /usr/include -L. -lcmdline -lcvip -lclassifier -llearutil -lboost_filesystem-gcc  -lblitz -L /usr/lib -lImlib2 -lfreetype -lz -L /usr/X11R6/lib -lX11 -lXext -ldl -lm

   结果：

   amadeuzou@ubuntu:~ /Projects/opencv/OLTbinaries/lib $ . /lib-detector person-1.jpg   model_4BiSVMLight.alt  result.jpg
      298    215    145    290    2.2674 1.10256009e-01
       13      9    237    475   3.71704 1.31164089e-01
      234     -7    230    460   3.59693 1.35627717e-01

3. 使用源码在learcode/app/中classify_rhog.cpp：配置及使用见learcode/README
4. 将模型数据copy到opencv的getDefaultPeopleDetector()，但要重新编译objdetect链接，没必要这么搞。从model中读取数据如下（共classifier->length() = 3780 + 1数据，1为classifier->linearbias）：

   int main( int argc, char ** argv) {
   
       string model_file = "model_4BiSVMLight.alt" ;
   
       LinearClassify* classifier = NULL;
   
       classifier = new LinearClassify(model_file, 0);
   
       cout<< "classifier->length() : " <<classifier->length()<<endl;
       FILE * fptr = fopen ( "hog.txt" , "w" );
           for ( int i = 0;i<classifier->length();i++){
               fprintf (fptr, "%1.8f " ,classifier->linearwt[i]);
               if (0 == (i+1) % 10) fprintf (fptr, "%s" , "\n" );
           }
       fprintf (fptr, "%1.8f " ,classifier->linearbias);
       fclose (fptr);
   
       return 0;
   
   }

5. opencv直接载入model: 将peopeldetect.cpp第49行hog.setSVMDetector(HOGDescriptor::getDefaultPeopleDetector())里面的vector换成你的模板数据，模板数据从model_4BiSVMLight.alt中载入，方法参照learcode/lib/windetect.cpp第1175行处LinearClassify::LinearClassify(std::string& modelfile, const int verbose)：

   /*
    * =====================================================================================
    *
    *       Filename:  model-detector.cpp
    *
    *    Description:  Code to detect object
    *
    *       Compiler:  gcc
    *
    *         Author:  Amadeu zou
    *         URL:  www.youtueye.com
    *
    * =====================================================================================
    */
   #include <fstream>
   #include <iostream>
   #include <vector>
   
   #include <stdio.h>
   #include <string.h>
   #include <ctype.h>
   
   #include "opencv2/imgproc/imgproc.hpp"
   #include "opencv2/objdetect/objdetect.hpp"
   #include "opencv2/highgui/highgui.hpp"
   
   using namespace std;
   using namespace cv;
   
   vector< float > load_lear_model( const char * model_file)
   {
        vector< float >  detector;
        FILE *modelfl;
        if ((modelfl = fopen (model_file, "rb" )) == NULL)
        {
            cout<< "Unable to open the modelfile" <<endl;
            return detector;
        }
   
        char version_buffer[10];
        if (! fread (&version_buffer, sizeof ( char ),10,modelfl))
        {
            cout<< "Unable to read version" <<endl;
            return