#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2\opencv.hpp>
#include <vector>
using namespace cv;
using namespace std;
int main()
{
CvCapture *capture = cvCreateFileCapture("bike.avi");
IplImage *img, *img_gray, *img_cny, *back, *fore;//原图像、灰度图像、边缘检测得到的图像、背景图像、前景图像
int width, height;//图像的长、宽
int cnt = 0; //视频帧数
img = cvQueryFrame(capture);
cnt++;
width = img->width; height = img->height;
vector< vector<double> > mean(width*height), sd(width*height), w(width*height);
img_gray = cvCreateImage(cvSize(width, height), IPL_DEPTH_8U, 1);
img_cny = cvCreateImage(cvSize(width, height), IPL_DEPTH_8U, 1);
back = cvCreateImage(cvSize(width, height), IPL_DEPTH_8U, 1);
fore = cvCreateImage(cvSize(width, height), IPL_DEPTH_8U, 1);
cvCvtColor(img, img_gray, CV_BGR2GRAY);
cvCopy(img_gray, back);
//初始化混合高斯背景模型
int K = 5, sd_init = 16, T = 100, match = 0;//高斯分布的个数、标准差的初始值 帧数阈值 匹配参数
double w_init = 0.05, D = 2.5, alph, p, threold = 0.7;;//权重的初始值、置信参数、学习速率、参数更新速率、预定阈值（用于选取背景模型）
for (int i = 0; i < height; i++)
{
for (int j = 0; j < width; j++)
{
mean[i*width + j].push_back((uchar)img_gray->imageData[i*width + j]);
sd[i*width + j].push_back(sd_init);
w[i*width + j].push_back(1);
}
}
while (1)
{
img = cvQueryFrame(capture);
if (!img) break;
cvCvtColor(img, img_gray, CV_BGR2GRAY);
cnt++;
//计算学习速率
if (cnt < T) alph = (double)1 / (2 * cnt);
else alph = (double)1 / (2 * T);
for (int i = 0; i < height; i++)
{
for (int j = 0; j < width; j++)
{
match = 0;
int len = mean[i*width + j].size();
for (int k = 0; k < len; k++)
{
//判断一个像素是否与背景模型匹配
if (fabs((uchar)img_gray->imageData[i*width + j] - mean[i*width + j][k]) < D*sd[i*width + j][k])
{
match = 1;
//更新权重、均值、标准差
w[i*width + j][k] = (1 - alph)*w[i*width + j][k] + alph;
p = alph / w[i*width + j][k];
mean[i*width + j][k] = (1 - p)*mean[i*width + j][k] + p*(uchar)img_gray->imageData[i*width + j];
sd[i*width + j][k] = sqrt((1 - p)*sd[i*width + j][k] * sd[i*width + j][k] + p*pow((uchar)img_gray->imageData[i*width + j] - mean[i*width + j][k], 2));
}
else
{
w[i*width + j][k] = (1 - alph)*w[i*width + j][k];
}
}
if (match == 0)
{
//增加新的背景模型
if (len < K)
{
mean[i*width + j].push_back((uchar)img_gray->imageData[i*width + j]);
sd[i*width + j].push_back(sd_init);
w[i*width + j].push_back(w_init);
}
//更改权重最小的背景模型
else
{
int min = 0;
double tmp = w[i*width + j][0];
for (int k = 1; k < len; k++)
{
if (tmp > w[i*width + j][k])
{
min = k;
tmp = w[i*width + j][k];
}
}
mean[i*width + j][min] = (uchar)img_gray->imageData[i*width + j];
sd[i*width + j][min] = sd_init;
w[i*width + j][min] = w_init;
}
}
//每隔50帧 删除权重最小的高斯分布
if (cnt % 50 == 0)
{
vector<double>::iterator it_w, it_sd, it_mean;
it_w = w[i*width + j].begin();
it_sd = sd[i*width + j].begin();
it_mean = mean[i*width + j].begin();
for (; it_w != w[i*width + j].end() && it_sd != sd[i*width + j].end() && it_mean != mean[i*width + j].end();)
{
if (*it_w < w_init && (*it_w / *it_sd) < (w_init / sd_init))
{
it_w = w[i*width + j].erase(it_w);
it_sd = sd[i*width + j].erase(it_sd);
it_mean = mean[i*width + j].erase(it_mean);
}
else
{
it_w ++;
it_sd ++;
it_mean ++;
}
}
}
//权重归一化
len = mean[i*width + j].size();
double tmp = 0;
for (int k = 0; k < len; k++)
{
tmp += w[i*width + j][k];
}
for (int k = 0; k < len; k++)
{
w[i*width + j][k] /= tmp;
}
//计算各背景模型的优先级，并以从大到小的顺序排序
double *rand;//优先级
int *rand_ind;//排序后的优先级索引
rand = (double *)malloc(sizeof(double)*len);
rand_ind = (int *)malloc(sizeof(int)*len);
for (int k = 0; k < len; k++)
{
rand[k] = w[i*width + j][k] / sd[i*width + j][k];
rand_ind[k] = k;
}
for (int k = 1; k< len; k++)
{
for (int l = 0; l < k; l++)
{
if (rand[k] > rand[l])
{
double rand_tmp = rand[k];
rand[k] = rand[l];
rand[l] = rand_tmp;
double rand_ind_tmp = rand_ind[k];
rand_ind[k] = rand_ind[l];
rand_ind[l] = rand_ind_tmp;
}
}
}
double sum = 0;
int B;
for (int k = 0; k < len; k++)
{
int temp = rand_ind[k];
sum += w[i*width + j][temp];
if (sum>threold)
{
B = k;
break;
}
}
match = 0;
back->imageData[i*width + j] = 0;
//求背景模型
for (int k = 0; k <= B; k++)
{
int temp = rand_ind[k];
back->imageData[i*width + j] += w[i*width + j][temp] * mean[i*width + j][temp];
}
/*
//求前景图像
for (int k = 0; k <= B; k++)
{
int temp = rand_ind[k];
if (fabs((uchar)img_gray->imageData[i*width + j] - mean[i*width + j][temp]) < D*sd[i*width + j][temp])
{
match = 1;
break;
}
}
if (match == 1) fore->imageData[i*width + j] = 0;
else fore->imageData[i*width + j] = (uchar)img_gray->imageData[i*width + j];
*/
free(rand);
free(rand_ind);
}
}
cvAbsDiff(img_gray, back, fore);
cvCanny(img_gray, img_cny, 1, 100);
cvShowImage("canny", img_cny);
cvAnd(img_cny, fore, fore);
cvShowImage("fore", fore);
cvShowImage("back", back);
cvShowImage("src", img);
char s = cvWaitKey(33);
if (s == 27) break;
}
cvDestroyAllWindows();
cvReleaseCapture(&capture);
cvReleaseImage(&img_gray);
cvReleaseImage(&fore);
cvReleaseImage(&back);
cvReleaseImage(&img_cny);
return 0;
}
评论0