OpenCv中混合高斯模型的实现

cvCreateGaussianBGModel///
CV_IMPL CvBGStatModel *cvCreateGaussianBGModel( IplImage*first_frame,CvGaussBGStatModelParams* parameters )
{
    CvGaussBGModel* bg_model = 0;
   
    CV_FUNCNAME( "cvCreateGaussianBGModel" );
   
    __BEGIN__;
   
    double var_init;
    CvGaussBGStatModelParams params;
    int i, j, k, m, n;
   
    // init parameters
    if( parameters == NULL )   
      {                       
        params.win_size      = CV_BGFG_MOG_WINDOW_SIZE;   // 初始化阶段的帧数；用户自定义模型学 习率a=1/win_size;
        params.bg_threshold = CV_BGFG_MOG_BACKGROUND_THRESHOLD;

        params.std_threshold = CV_BGFG_MOG_STD_THRESHOLD;   

        params.weight_init   = CV_BGFG_MOG_WEIGHT_INIT;

        params.variance_init = CV_BGFG_MOG_SIGMA_INIT*CV_BGFG_MOG_SIGMA_INIT; //方差
        params.minArea       = CV_BGFG_MOG_MINAREA;
        params.n_gauss       = CV_BGFG_MOG_NGAUSSIANS; //高斯分布函数的个数
    }
    else
    {
        params = *parameters; //用户自定义参数
    }
   
    if( !CV_IS_IMAGE(first_frame) )
        CV_ERROR( CV_StsBadArg, "Invalid or NULL first_frame parameter" );
   
    CV_CALL( bg_model = (CvGaussBGModel*)cvAlloc( sizeof(*bg_model) ));
    memset( bg_model, 0, sizeof(*bg_model) );
    bg_model->type = CV_BG_MODEL_MOG;    //CV_BG_MODEL_MOG为高斯背景模型
    bg_model->release = (CvReleaseBGStatModel)icvReleaseGaussianBGModel;
    bg_model->update = (CvUpdateBGStatModel)icvUpdateGaussianBGModel;
   
    bg_model->params = params;
   
    //prepare storages
    CV_CALL( bg_model->g_point = (CvGaussBGPoint*)cvAlloc(sizeof(CvGaussBGPoint)*
        ((first_frame->width*first_frame->height) + 256)));
   
    CV_CALL( bg_model->background = cvCreateImage(cvSize(first_frame->width,
        first_frame->height), IPL_DEPTH_8U, first_frame->nChannels));
    CV_CALL( bg_model->foreground = cvCreateImage(cvSize(first_frame->width,
        first_frame->height), IPL_DEPTH_8U, 1));
   
    CV_CALL( bg_model->storage = cvCreateMemStorage());
   
    //initializing
    var_init = 2 * params.std_threshold * params.std_threshold; //初始化方差
    CV_CALL( bg_model->g_point[0].g_values =
        (CvGaussBGValues*)cvAlloc( sizeof(CvGaussBGValues)*params.n_gauss*
        (first_frame->width*first_frame->height + 128)));
   
    for( i = 0, n = 0; i < first_frame->height; i++ ) //行
    {
        for( j = 0; j < first_frame->width; j++, n++ ) //列
        {
            const int p = i*first_frame->widthStep+j*first_frame->nChannels;
           //以下几步是对第一个高斯函数做初始化
            bg_model->g_point[n].g_values = bg_model->g_point[0].g_values + n*params.n_gauss;
            bg_model->g_point[n].g_values[0].weight = 1;    //权值赋为1
            bg_model->g_point[n].g_values[0].match_sum = 1; //高斯函数被匹配的次数
            for( m = 0; m < first_frame->nChannels; m++)
            {
                bg_model->g_point[n].g_values[0].variance[m] = var_init;

               //均值赋为当前像素的值
                bg_model->g_point[n].g_values[0].mean[m] = (unsigned char)first_frame->imageData[p + m];
            }

           //除第一以外的高斯分布函数的初始化(均值、权值和匹配次数都置零)
            for( k = 1; k < params.n_gauss; k++)  
            {   
                bg_model->g_point[n].g_values[k].weight = 0;
                bg_model->g_point[n].g_values[k].match_sum = 0;
                for( m = 0; m < first_frame->nChannels; m++){
                    bg_model->g_point[n].g_values[k].variance[m] = var_init;
                    bg_model->g_point[n].g_values[k].mean[m] = 0;
                }
            }
        }
    } //g_point[]:像素，g_values[]:高斯分布函数，mean[]:通道
   
    bg_model->countFrames = 0;
   
    __END__;
   
    if( cvGetErrStatus() < 0 )
    {
        CvBGStatModel* base_ptr = (CvBGStatModel*)bg_model;
       
        if( bg_model && bg_model->release )
            bg_model->release( &base_ptr );
        else
            cvFree( &bg_model );
        bg_model = 0;
    }
   
    return (CvBGStatModel*)bg_model;
}

 

cvUpdateBGStatModel(videoFrame,bgModel);
typedef int (CV_CDECL * CvUpdateBGStatModel)( IplImage* curr_frame, struct CvBGStatModel* bg_model );

/cvUpdateBGStatModel//
//函数功能：背景模型的更新，不仅要更新高斯分布函数的参数，还要更新各高斯函数的权重
static int CV_CDECL icvUpdateGaussianBGModel( IplImage* curr_frame, CvGaussBGModel* bg_model )
{
    int i, j, k, n;
    int region_count = 0;
    CvSeq *first_seq = NULL, *prev_seq = NULL, *seq = NULL;
   
    bg_model->countFrames++;
   
    for( i = 0, n = 0; i < curr_frame->height; i++ )
    {
        for( j = 0; j < curr_frame->width; j++, n++ )
        {
            int match[CV_BGFG_MOG_MAX_NGAUSSIANS];   //对高斯函数做标记，match[m]=1表示函数m为匹配的高斯分布函数
            double sort_key[CV_BGFG_MOG_MAX_NGAUSSIANS]; //此数组存贮每个高斯函数的均值与方差比值
            const int nChannels = curr_frame->nChannels;
            const int p = curr_frame->widthStep*i+j*nChannels;
            
           CvGaussBGPoint* g_point = &bg_model->g_point[n];
            const CvGaussBGStatModelParams bg_model_params = bg_model->params;
            double pixel[4];   // pixel[]存贮当前像素的各通道RGB值
            int no_match;
           
            for( k = 0; k < nChannels; k++ )
                pixel[k] = (uchar)curr_frame->imageData[p+k];
           
            no_match = icvMatchTest( pixel, nChannels, match, g_point, &bg_model_params ); //检查是否有与当前像素匹配的高斯函数
            if( bg_model->countFrames >= bg_model->params.win_size ) ？？？？？？？？？？？？？
            {
                icvUpdateFullWindow( pixel, nChannels, match, g_point, &bg_model->params );
                if( no_match == -1)
                    icvUpdateFullNoMatch( curr_frame, p, match, g_point, &bg_model_params );
            }
            else
            {
                icvUpdatePartialWindow( pixel, nChannels, match, g_point, &bg_model_params );
                if( no_match == -1)
                    icvUpdatePartialNoMatch( pixel, nChannels, match, g_point, &bg_model_params );
            }
            icvGetSortKey( nChannels, sort_key, g_point, &bg_model_params );
            icvInsertionSortGaussians( g_point, sort_key, (CvGaussBGStatModelParams *)&bg_model_params );
            icvBackgroundTest( nChannels, n, i, j, match, bg_model );
        }
    }
   
    //foreground filtering
   
    //filter small regions
    cvClearMemStorage(bg_model->storage);
   
    //cvMorphologyEx( bg_model->foreground, bg_model->foreground, 0, 0, CV_MOP_OPEN, 1 );
    //cvMorphologyEx( bg_model->foreground, bg_model->foreground, 0, 0, CV_MOP_CLOSE, 1 );
   
    cvFindContours( bg_model->foreground, bg_model->storage, &first_seq, sizeof(CvContour), CV_RETR_LIST );
    for( seq = first_seq; seq; seq = seq->h_next )
    {
        CvContour* cnt = (CvContour*)seq;
        if( cnt->rect.width * cnt->rect.height < bg_model->params.minArea )
        {
            //delete small contour
            prev_seq = seq->h_prev;
            if( prev_seq )
            {
                prev_seq->h_next = seq->h_next;
                if( seq->h_next ) seq->h_next->h_prev = prev_seq;
            }
            else
            {
                first_seq = seq->h_next;
                if( seq->h_next ) seq->h_next->h_prev = NULL;
            }
        }
        else
        {
            region_count++;
        }
    }
    bg_model->foreground_regions = first_seq;
    cvZero(bg_model->foreground);
    cvDrawContours(bg_model->foreground, first_seq, CV_RGB(0, 0, 255), CV_RGB(0, 0, 255), 10, -1);
   
    return region_count;
}

 

/icvMatchTest//
//函数功能：拿当前像素的值与已存在的高斯分布函数比较，查找是否存在匹配的的高斯分布函数，如果有则返回 k值(高斯分布函数的序号)
static int icvMatchTest( double* src_pixel, int nChannels, int* match,
                         const CvGaussBGPoint* g_point,
                         const CvGaussBGStatModelParams *bg_model_params )
{    //参数的传递：src_pixel为piexl[]:即当前像素的各通道值
    int k;
    int matchPosition=-1;
    for ( k = 0; k < bg_model_params->n_gauss; k++)
   match[k]=0;
   
    for ( k = 0; k < bg_model_params->n_gauss; k++)
    if (g_point->g_values[k].match_sum > 0)
{
        double sum_d2 = 0.0;
        double var_threshold = 0.0;
        for(int m = 0; m < nChannels; m++)
   {
            double d = g_point->g_values[k].mean[m]- src_pixel[m]; //通道m的原始模型值与当前像素的值之差
            sum_d2 += (d*d);
            var_threshold += g_point->g_values[k].variance[m];
        }  

      
       //当前sum_d2为d0,d1,d2的平方和，var_threshold的值为像素各通道方差之和
        var_threshold = bg_model_params->std_threshold*

                                 bg_model_params- >std_threshold*var_threshold;
        if(sum_d2 < var_threshold) //查看是否可以与某高斯分布匹配 ？？？？？？？？？？？？？？？？
   {
            match[k] = 1;
            matchPosition = k;
            break; //如果和第k个高斯函数匹配，则终止与后续函数的匹配
        }
    }
   
    return matchPosition;
}

 

///icvUpdateFullWindow
//函数功能：更新每个高斯分布的权值(对匹配的高斯函数k加大权值，其余的则减小权值)，如果前面的结果中存在匹配的高斯分布函数k,则需要再对第k个高斯分布函数的均值mean和方差variance做修正
static void icvUpdateFullWindow( double* src_pixel, int nChannels, int* match,
                                 CvGaussBGPoint* g_point,
                                 const CvGaussBGStatModelParams *bg_model_params )
{ //参数的传递：src_pixel为piexl[]:即当前帧中该像素的RGB值
    const double learning_rate_weight = (1.0/(double)bg_model_params->win_size); //用户自定义模型学习率a

    for(int k = 0; k < bg_model_params->n_gauss; k++)
{  
   //对每个高斯分布的权值做修正：w=(1-a)w+a*m (a:模型学习率，m是匹配，匹配就是1，不匹配就是0)
        g_point->g_values[k].weight = g_point->g_values[k].weight +
            (learning_rate_weight*((double)match[k] -g_point->g_values[k].weight));

        if(match[k]) //如果存在匹配的高斯分布函数k（当前像素为背景像素）,则需要再对第k个高斯分布函数的均值mean和方差variance更新
   {
            double learning_rate_gaussian = (double)match[k]/(g_point->g_values[k].weight*
                (double)bg_model_params->win_size); //参数学习率p(p=a/w)
            for(int m = 0; m < nChannels; m++)
    {   //参数更新公式：u=(1-p)*u0+p*x; o*o=(1-p)*o*o+p*tmpDiff*tmpDiff
                const double tmpDiff = src_pixel[m] - g_point->g_values[k].mean[m]; //当前像素的通道m的值与原始模型值之差
                g_point->g_values[k].mean[m] = g_point->g_values[k].mean[m] + (learning_rate_gaussian * tmpDiff);
                g_point->g_values[k].variance[m] = g_point->g_values[k].variance[m]+
                    (learning_rate_gaussian*((tmpDiff*tmpDiff) - g_point->g_values[k].variance[m]));
            }
        }
    }
}

 

 

 

/icvUpdatePartialWindow/
//函数功能：对所有的高斯分布函数做更新.至少每个高斯分布的权值必须修正，如果前面的结果中存在匹配的高斯分布函数k,则需要再对第k个高斯分布函数的match_sum修改，最终对那些匹配的高斯分布函数k的参数match_sum>0的做均值mean和方差variance修正
static void icvUpdatePartialWindow( double* src_pixel, int nChannels, int* match, CvGaussBGPoint* g_point, const CvGaussBGStatModelParams *bg_model_params )
{
    int k, m;
    int window_current = 0;
   
    for( k = 0; k < bg_model_params->n_gauss; k++ )
        window_current += g_point->g_values[k].match_sum;   //window_current为k个高斯分布函数的match_sum值之和
   
    for( k = 0; k < bg_model_params->n_gauss; k++ )
    {
        g_point->g_values[k].match_sum += match[k]; //修正匹配的高斯分布函数k的match_sum值
        double learning_rate_weight = (1.0/((double)window_current + 1.0)); //increased by one since sum
         //修正每个高斯分布的权值
        g_point->g_values[k].weight = g_point->g_values[k].weight +
            (learning_rate_weight*((double)match[k] - g_point->g_values[k].weight));
       
        if( g_point->g_values[k].match_sum > 0 && match[k] )
        {
            double learning_rate_gaussian = (double)match[k]/((double)g_point->g_values[k].match_sum);
            for( m = 0; m < nChannels; m++ )
            {
                const double tmpDiff = src_pixel[m] - g_point->g_values[k].mean[m];
                g_point->g_values[k].mean[m] = g_point->g_values[k].mean[m] +
                    (learning_rate_gaussian*tmpDiff);
                g_point->g_values[k].variance[m] = g_point->g_values[k].variance[m]+
                    (learning_rate_gaussian*((tmpDiff*tmpDiff) - g_point->g_values[k].variance[m]));
            }
        }
    }
}

 

//icvUpdateFullNoMatch//
//函数功能：当所有的高斯函数均不匹配时，说明有新的分布出现，需要将原高斯函数中sort_key最小的替换为新的高斯函数(权值小，方差大),其余的高斯函数对应的只需更新权值
static void icvUpdateFullNoMatch( IplImage* gm_image, int p, int* match,
                                  CvGaussBGPoint* g_point,
                                  const CvGaussBGStatModelParams *bg_model_params)
{ //参数的传递：gm_image为当前帧curr_frame
    int k, m;
    double alpha;
    int match_sum_total = 0;

    g_point->g_values[bg_model_params->n_gauss - 1].match_sum = 1;    //将新的高斯分布函数的match_sum置为1
   
   
    for( k = 0; k < bg_model_params->n_gauss ; k++ )
        match_sum_total += g_point->g_values[k].match_sum;

    g_point->g_values[bg_model_params->n_gauss - 1].weight = 1./(double)match_sum_total; //要给新的高斯分布函数赋一个较小的权值

    //将新的高斯分布函数的variance[m]全部置为variance_init；mean[m]的值置为当前像素各通道的值
    for( m = 0; m < gm_image->nChannels ; m++ )
    {
        g_point->g_values[bg_model_params->n_gauss - 1].variance[m] = bg_model_params->variance_init;
        g_point->g_values[bg_model_params->n_gauss - 1].mean[m] = (unsigned char)gm_image->imageData[p + m];
    }
    
//对其他的高斯分布函数做权值更新：w=(1-a)*w+a*m (a:模型学习率，m是匹配，匹配就是1，不匹配就是0)
    alpha = 1.0 - (1.0/bg_model_params->win_size);   //alpha=1-a;
    for( k = 0; k < bg_model_params->n_gauss - 1; k++ )
    {  
        g_point->g_values[k].weight *= alpha;
        if( match[k] )
            g_point->g_values[k].weight += alpha;
    }
}

 

 

icvUpdatePartialNoMatch
static void
icvUpdatePartialNoMatch(double *pixel,
                        int nChannels,
                        int* /*match*/,
                        CvGaussBGPoint* g_point,
                        const CvGaussBGStatModelParams *bg_model_params)
{
    int k, m;
    //new value of last one
    g_point->g_values[bg_model_params->n_gauss - 1].match_sum = 1;
   
    //get sum of all but last value of match_sum
    int match_sum_total = 0;
    for(k = 0; k < bg_model_params->n_gauss ; k++)
        match_sum_total += g_point->g_values[k].match_sum;

    for(m = 0; m < nChannels; m++)
    {
        //first pass mean is image value
        g_point->g_values[bg_model_params->n_gauss - 1].variance[m] = bg_model_params->variance_init;
        g_point->g_values[bg_model_params->n_gauss - 1].mean[m] = pixel[m];
    }
    for(k = 0; k < bg_model_params->n_gauss; k++)
    { //更新所有高斯分布的权值
        g_point->g_values[k].weight = (double)g_point->g_values[k].match_sum /
            (double)match_sum_total;
    }
}

 

/icvGetSortKey///
//函数功能：计算各个高斯分布weight/sqrt(variance_sum)的值，后面将对该值进行排序（该值越大则表示背景的可能性就越大）
static void icvGetSortKey( const int nChannels, double* sort_key, const CvGaussBGPoint* g_point,
                           const CvGaussBGStatModelParams *bg_model_params )
{
    int k, m;
    for( k = 0; k < bg_model_params->n_gauss; k++ )
    {
        // Avoid division by zero
        if( g_point->g_values[k].match_sum > 0 )
        {
            // Independence assumption between components
            double variance_sum = 0.0;
            for( m = 0; m < nChannels; m++ )
                variance_sum += g_point->g_values[k].variance[m];
           
            sort_key[k] = g_point->g_values[k].weight/sqrt(variance_sum);     //sort_key=w/(o*o)
        }
        else
            sort_key[k]= 0.0;
    }
}

 

//icvInsertionSortGaussians
static void icvInsertionSortGaussians( CvGaussBGPoint* g_point, double* sort_key, CvGaussBGStatModelParams *bg_model_params )
{
    int i, j;
    for( i = 1; i < bg_model_params->n_gauss; i++ )
    {
        double index = sort_key[i];
        for( j = i; j > 0 && sort_key[j-1] < index; j-- )   //对sort_key[]按降序排序
        {
            double temp_sort_key = sort_key[j];
            sort_key[j] = sort_key[j-1];
            sort_key[j-1] = temp_sort_key;
           
            CvGaussBGValues temp_gauss_values = g_point->g_values[j];
            g_point->g_values[j] = g_point->g_values[j-1];
            g_point->g_values[j-1] = temp_gauss_values;
        }
//        sort_key[j] = index;
    }
}

 

///icvBackgroundTest/
static void icvBackgroundTest( const int nChannels, int n, int i, int j, int *match, CvGaussBGModel* bg_model )
{
    int m, b;
    uchar pixelValue = (uchar)255;   // 像素默认都为前景
    double weight_sum = 0.0;
    CvGaussBGPoint* g_point = bg_model->g_point;
   
    for( m = 0; m < nChannels; m++)？？？？？？？？？？？？？
        bg_model->background->imageData[ bg_model->background->widthStep*i + j*nChannels + m] = (unsigned char)(g_point[n].g_values[0].mean[m]+0.5);
   
    for( b = 0; b < bg_model->params.n_gauss; b++)
    {
        weight_sum += g_point[n].g_values[b].weight;
        if( match[b] )
            pixelValue = 0;   //if为真，说明该像素已与某高斯函数匹配，该像素为背景
        if( weight_sum > bg_model->params.bg_threshold )  
            break; //如果if语句为真，则前b个高斯分布被选为描述背景的函数
    }
   
    bg_model->foreground->imageData[ bg_model->foreground->widthStep*i + j] = pixelValue;
}

转载于:https://www.cnblogs.com/hjqxaly/archive/2010/10/07/1845069.html