http://www.cnblogs.com/tornadomeet/archive/2012/06/02/2531565.html
里面有了GMM的代码，并有了详细的注释。我之前根据这个代码（在这里，非常感谢tornadomeet）改写了一个Mat格式的版本，现在发上来和大家交流，具体如下：（在VS2010+OpenCV2.4.2中测试通过）。（当然了，OpenCV也已经提供了MOG的背景减除方法）
MOG_BGS.h
[cpp] view plain copy #pragma once #include <iostream> #include "opencv2/opencv.hpp" using namespace cv; using namespace std; //定义gmm模型用到的变量 #define GMM_MAX_COMPONT 6 //每个GMM最多的高斯模型个数 #define GMM_LEARN_ALPHA 0.005 #define GMM_THRESHOD_SUMW 0.7 #define TRAIN_FRAMES 60 // 对前 TRAIN_FRAMES 帧建模 class MOG_BGS { public: MOG_BGS(void); ~MOG_BGS(void); void init(const Mat _image); void processFirstFrame(const Mat _image); void trainGMM(const Mat _image); void getFitNum(const Mat _image); void testGMM(const Mat _image); Mat getMask(void){return m_mask;}; private: Mat m_weight[GMM_MAX_COMPONT]; //权值 Mat m_mean[GMM_MAX_COMPONT]; //均值 Mat m_sigma[GMM_MAX_COMPONT]; //方差 Mat m_mask; Mat m_fit_num; };
MOG_BGS.cpp
[cpp] view plain copy #include "MOG_BGS.h" MOG_BGS::MOG_BGS(void) { } MOG_BGS::~MOG_BGS(void) { } // 全部初始化为0 void MOG_BGS::init(const Mat _image) { /****initialization the three parameters ****/ for(int i = 0; i < GMM_MAX_COMPONT; i++) { m_weight[i] = Mat::zeros(_image.size(), CV_32FC1); m_mean[i] = Mat::zeros(_image.size(), CV_8UC1); m_sigma[i] = Mat::zeros(_image.size(), CV_32FC1); } m_mask = Mat::zeros(_image.size(),CV_8UC1); m_fit_num = Mat::ones(_image.size(),CV_8UC1); } //gmm第一帧初始化函数实现 //捕获到第一帧时对高斯分布进行初始化．主要包括对每个高斯分布的权值、期望和方差赋初值． //其中第一个高斯分布的权值为1，期望为第一个像素数据．其余高斯分布权值为0，期望为0． //每个高斯分布都被赋予适当的相等的初始方差 15 void MOG_BGS::processFirstFrame(const Mat _image) { for(int i = 0; i < GMM_MAX_COMPONT; i++) { if (i == 0) { m_weight[i].setTo(1.0); _image.copyTo(m_mean[i]); m_sigma[i].setTo(15.0); } else { m_weight[i].setTo(0.0); m_mean[i].setTo(0); m_sigma[i].setTo(15.0); } } } // 通过新的帧来训练GMM void MOG_BGS::trainGMM(const Mat _image) { for(int i = 0; i < _image.rows; i++) { for(int j = 0; j < _image.cols; j++) { int num_fit = 0; /**************************** Update parameters Start ******************************************/ for(int k = 0 ; k < GMM_MAX_COMPONT; k++) { int delm = abs(_image.at<uchar>(i, j) - m_mean[k].at<uchar>(i, j)); long dist = delm * delm; // 判断是否匹配：采样值与高斯分布的均值的距离小于3倍方差（表示匹配） if( dist < 3.0 * m_sigma[k].at<float>(i, j)) { // 如果匹配 /****update the weight****/ m_weight[k].at<float>(i, j) += GMM_LEARN_ALPHA * (1 - m_weight[k].at<float>(i, j)); /****update the average****/ m_mean[k].at<uchar>(i, j) += (GMM_LEARN_ALPHA / m_weight[k].at<uchar>(i, j)) * delm; /****update the variance****/ m_sigma[k].at<float>(i, j) += (GMM_LEARN_ALPHA / m_weight[k].at<float>(i, j)) * (dist - m_sigma[k].at<float>(i, j)); } else { // 如果不匹配。则该该高斯模型的权值变小 m_weight[k].at<float>(i, j) += GMM_LEARN_ALPHA * (0 - m_weight[k].at<float>(i, j)); num_fit++; // 不匹配的模型个数 } } /**************************** Update parameters End ******************************************/ /*********************** Sort Gaussian component by 'weight / sigma' Start ****************************/ //对gmm各个高斯进行排序,从大到小排序,排序依据为 weight / sigma for(int kk = 0; kk < GMM_MAX_COMPONT; kk++) { for(int rr=kk; rr< GMM_MAX_COMPONT; rr++) { if(m_weight[rr].at<float>(i, j)/m_sigma[rr].at<float>(i, j) > m_weight[kk].at<float>(i, j)/m_sigma[kk].at<float>(i, j)) { //权值交换 float temp_weight = m_weight[rr].at<float>(i, j); m_weight[rr].at<float>(i, j) = m_weight[kk].at<float>(i, j); m_weight[kk].at<float>(i, j) = temp_weight; //均值交换 uchar temp_mean = m_mean[rr].at<uchar>(i, j); m_mean[rr].at<uchar>(i, j) = m_mean[kk].at<uchar>(i, j); m_mean[kk].at<uchar>(i, j) = temp_mean; //方差交换 float temp_sigma = m_sigma[rr].at<float>(i, j); m_sigma[rr].at<float>(i, j) = m_sigma[kk].at<float>(i, j); m_sigma[kk].at<float>(i, j) = temp_sigma; } } } /*********************** Sort Gaussian model by 'weight / sigma' End ****************************/ /*********************** Create new Gaussian component Start ****************************/ if(num_fit == GMM_MAX_COMPONT && 0 == m_weight[GMM_MAX_COMPONT - 1].at<float>(i, j)) { //if there is no exit component fit,then start a new component //当有新值出现的时候，若目前分布个数小于M，新添一个分布，以新采样值作为均值，并赋予较大方差和较小权值 for(int k = 0 ; k < GMM_MAX_COMPONT; k++) { if(0 == m_weight[k].at<float>(i, j)) { m_weight[k].at<float>(i, j) = GMM_LEARN_ALPHA; m_mean[k].at<uchar>(i, j) = _image.at<uchar>(i, j); m_sigma[k].at<float>(i, j) = 15.0; //normalization the weight,let they sum to 1 for(int q = 0; q < GMM_MAX_COMPONT && q != k; q++) { //对其他的高斯模型的权值进行更新，保持权值和为1 /****update the other unfit's weight,u and sigma remain unchanged****/ m_weight[q].at<float>(i, j) *= (1 - GMM_LEARN_ALPHA); } break; //找到第一个权值不为0的即可 } } } else if(num_fit == GMM_MAX_COMPONT && m_weight[GMM_MAX_COMPONT -1].at<float>(i, j) != 0) { //如果GMM_MAX_COMPONT都曾经赋值过，则用新来的高斯代替权值最弱的高斯，权值不变，只更新均值和方差 m_mean[GMM_MAX_COMPONT-1].at<uchar>(i, j) = _image.at<uchar>(i, j); m_sigma[GMM_MAX_COMPONT-1].at<float>(i, j) = 15.0; } /*********************** Create new Gaussian component End ****************************/ } } } //对输入图像每个像素gmm选择合适的高斯分量个数 //排序后最有可能是背景分布的排在最前面，较小可能的短暂的分布趋向于末端．我们将排序后的前fit_num个分布选为背景模型; //在排过序的分布中，累积概率超过GMM_THRESHOD_SUMW的前fit_num个分布被当作背景模型，剩余的其它分布被当作前景模型． void MOG_BGS::getFitNum(const Mat _image) { for(int i = 0; i < _image.rows; i++) { for(int j = 0; j < _image.cols; j++) { float sum_w = 0.0; //重新赋值为0，给下一个像素做累积 for(uchar k = 0; k < GMM_MAX_COMPONT; k++) { sum_w += m_weight[k].at<float>(i, j); if(sum_w >= GMM_THRESHOD_SUMW) //如果这里THRESHOD_SUMW=0.6的话，那么得到的高斯数目都为1，因为每个像素都有一个权值接近1 { m_fit_num.at<uchar>(i, j) = k + 1; break; } } } } } //gmm测试函数的实现 void MOG_BGS::testGMM(const Mat _image) { for(int i = 0; i < _image.rows; i++) { for(int j = 0; j < _image.cols; j++) { int k = 0; for( ; k < m_fit_num.at<uchar>(i, j); k++) { if(abs(_image.at<uchar>(i, j) - m_mean[k].at<uchar>(i, j)) < (uchar)( 2.5 * m_sigma[k].at<float>(i, j))) { m_mask.at<uchar>(i, j) = 0; break; } } if(k == m_fit_num.at<uchar>(i, j)) { m_mask.at<uchar>(i, j) = 255; } } } }
Main.cpp
[cpp] view plain copy // This is based on the "An Improved Adaptive Background Mixture Model for // Real-time Tracking with Shadow Detection" by P. KaewTraKulPong and R. Bowden // Author : zouxy // Date : 2013-4-13 // HomePage : http://blog.csdn.net/zouxy09 // Email : zouxy09@qq.com #include "opencv2/opencv.hpp" #include "MOG_BGS.h" #include <iostream> #include <cstdio> using namespace cv; using namespace std; int main(int argc, char* argv[]) { Mat frame, gray, mask; VideoCapture capture; capture.open("video.avi"); if (!capture.isOpened()) { cout<<"No camera or video input!\n"<<endl; return -1; } MOG_BGS Mog_Bgs; int count = 0; while (1) { count++; capture >> frame; if (frame.empty()) break; cvtColor(frame, gray, CV_RGB2GRAY); if (count == 1) { Mog_Bgs.init(gray); Mog_Bgs.processFirstFrame(gray); cout<<" Using "<<TRAIN_FRAMES<<" frames to training GMM..."<<endl; } else if (count < TRAIN_FRAMES) { Mog_Bgs.trainGMM(gray); } else if (count == TRAIN_FRAMES) { Mog_Bgs.getFitNum(gray); cout<<" Training GMM complete!"<<endl; } else { Mog_Bgs.testGMM(gray); mask = Mog_Bgs.getMask(); morphologyEx(mask, mask, MORPH_OPEN, Mat()); erode(mask, mask, Mat(7, 7, CV_8UC1), Point(-1, -1)); // You can use Mat(5, 5, CV_8UC1) here for less distortion dilate(mask, mask, Mat(7, 7, CV_8UC1), Point(-1, -1)); imshow("mask", mask); } imshow("input", frame); if ( cvWaitKey(10) == 'q' ) break; } return 0; }
原文地址：http://blog.csdn.net/zouxy09/article/details/9622401
