OpenCV HOG PeopleDetector：从训练到检测 1

转载：http://blog.sciencenet.cn/blog-261330-548546.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：

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

71

72

73

74

75

76

77

78

79

80

81

82

83

84

85

86

87

88

89

90

91

92

93

94

95

96

97

98

99

100

101

102

103

104

105

106

107

108

109

110

111

#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

1

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：

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

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：

1

2

3

4

5

6

7

8

9

10

11

12

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！