近来,有不少人咨询我关于VIBE算法的问题,而且对于有些细节问题懵懵懂懂,索要源码类的,考虑这个算法的应用以及很多人对此有比较深的兴趣,遂将其放在博客上供大家学习。该版本的代码是在学校的时候写的,里面也加入了一些其他的后处理内容,尽管还有不足,但是对于加深对VIBE算法的理解肯定有一定帮助。另外,由于最新版本的代码在公司电脑上,不便提供。关于理论方面的请参考下面二篇文章:
1)VIBE-A powerful random technique to estimatie the background in video sequences.
2) VIBE-A universal background subtraction algorithms for video sequences
VIBE的头文件Vibe.hpp如下:
#pragma once
#include "stdafx.h"
#define WINSIZE 3
class Vibe
{
public:
Vibe(void);
Vibe(IplImage *img);
void SetMinMatch(int nthreshold){g_MinMatch=nthreshold;}
void SetRadius(int radius){g_Radius=radius;}
void SetSampleNum(int num){g_SampleNum=num;}
void SetThreshold(double t){g_threshold=t;}
IplImage* GetForeground(){return g_ForeImg;}
IplImage* GetSegMask(){return g_SegementMask;}
void Detect(IplImage *img);
void ForegroundCombineEdge(); // 结合边缘信息
void DeleteSmallAreaInForeground(double minArea=20);//删除小面积区域
// 实现背景更新机制
void Update();
// 实现后处理,主要用形态学算子
void PostProcess();
public:
~Vibe(void);
private:
void ClearLongLifeForeground(int i_lifeLength=200); // 清除场景中存在时间较长的像素,i_lifeLength用于控制允许存在的最长时间
double AreaDense(IplImage *pFr,int AI,int AJ,int W,int H); //计算(i,j)处邻域大小为W×H的密度
int GetRandom(int istart,int iend); // 默认istart=0,iend=15
int GetRandom(int random);
int GetRandom();// 产生一个随机数
// 计算两个像素之间的欧式距离
double CalcPixelDist(CvScalar bkCs,CvScalar curCs);
// 按照Kim的方法来计算颜色畸变
double CalcuColorDist(CvScalar bkCs,CvScalar curCs);
int g_SampleNum;// Sample number for the models,默认为20
int g_MinMatch; // 当前像素与背景模型匹配的最少个数,默认为2
int g_Height;
int g_Width;
int g_Radius;// 球体的半径,默认为20
int g_offset; //边界的宽和高
double g_threshold; // 距离度量的阈值
unsigned char ***g_Model;// 保存背景模型
IplImage *g_ForeImg;// 保存前景图
IplImage *g_Edge;
IplConvKernel* element;
IplImage *g_SegementMask; //分割掩膜
IplImage *g_UpdateMask; // 更新掩膜
IplImage *g_Gray;
int ** LifeLength; // 记录前景点的生命长度,如果前景点的生命长度到达一定的阈值,则将其融入背景中去,且要随机两次。
};
对应的实现文件如下Vibe.cpp所示:
#include "StdAfx.h"
#include "Vibe.h"
Vibe::Vibe(void)
{
g_Radius=20;
g_MinMatch=2;
g_SampleNum=20;
g_offset=(WINSIZE-1)/2;
}
Vibe::Vibe(IplImage *img)
{
if (!img)
{
cout<<" The parameter referenced to NUll Pointer!"<<endl;
return;
}
this->g_Height=img->height;
this->g_Width=img->width;
g_Radius=20;
g_MinMatch=2;
g_SampleNum=20;
g_threshold=50;
g_offset=(WINSIZE-1)/2;
g_ForeImg=cvCreateImage(cvGetSize(img),IPL_DEPTH_8U,1);
g_Gray=cvCreateImage(cvGetSize(img),IPL_DEPTH_8U,1);
g_Edge=cvCreateImage(cvGetSize(img),IPL_DEPTH_8U,1);
g_SegementMask=cvCreateImage(cvGetSize(img),IPL_DEPTH_8U,1);
g_UpdateMask=cvCreateImage(cvGetSize(img),IPL_DEPTH_8U,1);
element=cvCreateStructuringElementEx(3,3,1,1,CV_SHAPE_CROSS,NULL);
cvCvtColor(img,g_Gray,CV_BGR2GRAY);
// 以上完成相关的初始化操作
/********************** 以下实现第一帧在每个像素的8邻域内的采样功能,建立对应的背景模型*****************************/
int i=0,j=0,k=0;
g_Model=new unsigned char**[g_SampleNum];
for (k=0;k<g_SampleNum;k++)
{
g_Model[k]=new unsigned char *[g_Height];
for(i=0;i<g_Height;i++)
{
g_Model[k][i]=new unsigned char [g_Width];
for (j=0;j<g_Width;j++)
{
g_Model[k][i][j]=0;
}
}
}
// 采样进行背景建模
double dVal;
int ri=0,rj=0; //随机采样的值
for (i=g_offset;i<g_Height-g_offset;i++)
{
for (j=g_offset;j<g_Width-g_offset;j++)
{
// 周围3*3的邻域内进行采样
for(k=0;k<g_SampleNum;k++)
{
ri=GetRandom(i);
rj=GetRandom(j);
dVal=cvGetReal2D(g_Gray,ri,rj);
g_Model[k][i][j]=dVal;
}
}
}
// 初始化前景点掩膜的生命长度
LifeLength=new int *[g_Height];
for (i=0;i<g_Height;i++)
{
LifeLength[i]=new int [g_Width];
for(j=0;j<g_Width;j++)
{
LifeLength[i][j]=0;
}
}
}
void Vibe::Detect(IplImage *img)
{
cvZero(g_ForeImg);
cvCvtColor(img,g_Gray,CV_BGR2GRAY);
int i=0,j=0,k=0;
double dModVal,dCurrVal;
int tmpCount=0;// 距离比较在阈值内的次数
double tmpDist=0;
int iR1,iR2;//产生随机数
int Ri,Rj; // 产生邻域内X和Y的随机数
for (i=0;i<g_Height;i++)
{
for (j=0;j<g_Width;j++)
{
if( i < g_offset || j < g_offset || i> g_Height - g_offset || j> g_Width - g_offset )
{
cvSetReal2D(g_ForeImg,i,j,0);
continue;
}
else
{
tmpCount=0;
dCurrVal=cvGetReal2D(g_Gray,i,j);
for (k=0;k<g_SampleNum && tmpCount<g_MinMatch ;k++)
{
dModVal=g_Model[k][i][j];
//tmpDist=CalcPixelDist(dCurrVal,dModVal);
//tmpDist=CalcuColorDist(dCurrVal,dModVal);
tmpDist=fabs(dModVal-dCurrVal);
if (tmpDist<g_Radius)
{
tmpCount++;
}
}
//判断是否匹配上
if (tmpCount>=g_MinMatch)
{
cvSetReal2D(g_ForeImg,i,j,0);
// 背景模型的更新
iR1=GetRandom(0,15);
if (iR1==0)
{
iR2=GetRandom();
g_Model[iR2][i][j]=dCurrVal;
}
//进一步更新邻域模型
iR1=GetRandom(0,15);
if (iR1==0)
{
Ri=GetRandom(i);
Rj=GetRandom(j);
iR2=GetRandom();
g_Model[iR2][Ri][Rj]=dCurrVal;
}
}
else
{
cvSetReal2D(g_ForeImg,i,j,255);
}
}
}
}
//ForegroundCombineEdge();
DeleteSmallAreaInForeground(80);
ClearLongLifeForeground();
//PostProcess();
}
double Vibe::AreaDense(IplImage *pFr,int AI,int AJ,int W,int H)
{
if (AI<=2 || AJ<=2 || AJ>=(g_Width-2) || AI>=(g_Height-2))
{
return 0;
}
int Num=0,i=0,j=0;
double dVal=0,dense=0;
int Total=(2*H+1)*(2*W+1);
for (i=AI-H;i<=AI+H;i++)
{
for (j=AJ-W;j<=AJ+W;j++)
{
dVal=cvGetReal2D(pFr,i,j);
if (dVal>200)
{
Num++;
}
}
}
dense=(double)Num/(double)Total;
return dense;
}
void Vibe::ForegroundCombineEdge()
{
cvZero(g_Edge);
//cvZero(g_SegementMask);
//cvCopy(g_ForeImg,g_SegementMask);
cvCanny(g_Gray,g_Edge,30,200,3);
int i=0,j=0;
double dense;
double dVal;
for (i=g_offset;i<g_Height-g_offset;i++)
{
for (j=g_offset;j<g_Width-g_offset;j++)
{
dense=AreaDense(g_ForeImg,i,j,2,2);
dVal=cvGetReal2D(g_Edge,i,j);
if (dense>0.2 && dVal>200)
{
cvSetReal2D(g_ForeImg,i,j,255);
}
}
}
}
void Vibe::DeleteSmallAreaInForeground(double minArea/* =20 */)
{
//cvZero(g_SegementMask);
//cvCopy(g_ForeImg,g_SegementMask);
int region_count = 0;
CvSeq *first_seq = NULL, *prev_seq = NULL, *seq = NULL;
CvMemStorage* storage = cvCreateMemStorage();
cvClearMemStorage(storage);
cvFindContours( g_ForeImg, 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 < minArea )
{
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++;
}
}
cvZero(g_ForeImg);
cvDrawContours(g_ForeImg, first_seq, CV_RGB(0, 0, 255), CV_RGB(0, 0, 255), 10, -1);
/*
CvContourScanner scanner = cvStartFindContours( g_ForeImg, storage,sizeof(CvContour), CV_RETR_EXTERNAL, CV_CHAIN_APPROX_SIMPLE, cvPoint(0,0) );
CvSeq *contours=NULL,*c=NULL;
int poly1Hull0=0;
int nContours=0;
double perimScale=100;
while( (c = cvFindNextContour( scanner )) != 0 )
{
double len = cvContourPerimeter( c );
double q = (g_ForeImg->height + g_ForeImg->width)/perimScale; // calculate perimeter len threshold
if( len < q ) //Get rid of blob if it's perimeter is too small
cvSubstituteContour( scanner, 0 );
else //Smooth it's edges if it's large enough
{
CvSeq* newC;
if( poly1Hull0 ) //Polygonal approximation of the segmentation
newC = cvApproxPoly( c, sizeof(CvContour), storage, CV_POLY_APPROX_DP, 2, 0 );
else //Convex Hull of the segmentation
newC = cvConvexHull2( c, storage, CV_CLOCKWISE, 1 );
cvSubstituteContour( scanner, newC );
nContours++;
}
}
contours = cvEndFindContours( &scanner );
// paint the found regions back into the image
cvZero( g_ForeImg );
for( c=contours; c != 0; c = c->h_next )
cvDrawContours( g_ForeImg, c, cvScalarAll(255), cvScalarAll(0), -1, CV_FILLED, 8,cvPoint(0,0));
*/
cvReleaseMemStorage(&storage);
}
void Vibe::ClearLongLifeForeground(int i_lifeLength/* =200 */)
{
int i=0,j=0;
double dVal=0;
double dLife=0;
int iR1,iR2=0;
double dCurrVal=0;
for (i=g_offset;i<g_Height-g_offset;i++)
{
for (j=g_offset;j<g_Width-g_offset;j++)
{
dVal=cvGetReal2D(g_ForeImg,i,j);
dLife=LifeLength[i][j];
if (dLife>i_lifeLength)
{
LifeLength[i][j]=0;
dCurrVal=cvGetReal2D(g_Gray,i,j);
// 更新背景模型
iR1=GetRandom();
iR2=GetRandom();
g_Model[iR1][i][j]=dCurrVal;
g_Model[iR2][i][j]=dCurrVal;
}
else
{
LifeLength[i][j]=dLife+1;
}
}
}
}
void Vibe::Update()
{
cvZero(g_UpdateMask);
}
void Vibe::PostProcess()
{
cvZero(g_SegementMask);
cvMorphologyEx(g_ForeImg,g_SegementMask,NULL,element,CV_MOP_OPEN,1);
}
//算颜色畸变
double Vibe::CalcuColorDist(CvScalar bkCs,CvScalar curCs)
{
double r,g,b,br,bg,bb;
r=curCs.val[0];
g=curCs.val[1];
b=curCs.val[2];
br=bkCs.val[0];
bg=bkCs.val[1];
bb=bkCs.val[2];
double curDist=r*r+g*g*b*b;
double bkDist=br*br+bg*bg+bb*bb;
double curBK=r*br+g*bg+b*bb;
double curbkDist=curBK*curBK;
double SquareP;
if (bkDist==0.0)
{
SquareP=0;
}
else
{
SquareP=curbkDist/bkDist;
}
double dist=sqrtf(curDist-SquareP);
return dist;
}
double Vibe::CalcPixelDist(CvScalar bkCs,CvScalar curCs)
{
double tmpDist=pow(bkCs.val[0]-curCs.val[0],2)+pow(bkCs.val[1]-curCs.val[1],2)+pow(bkCs.val[2]-curCs.val[2],2);
return sqrtf(tmpDist);
}
int Vibe::GetRandom()
{
int val = g_SampleNum * 1.0 * rand() / RAND_MAX;
if( val == g_SampleNum )
return val - 1;
else
return val;
}
int Vibe::GetRandom(int random)
{
int val=random-g_offset+rand()%(2*g_offset);
if (val<random-g_offset)
{
val=random-g_offset;
}
if (val>random+g_offset)
{
val=random+g_offset;
}
return val;
}
int Vibe::GetRandom(int istart,int iend)
{
int val=istart+rand()%(iend-istart);
return val;
}
Vibe::~Vibe(void)
{
if (g_ForeImg)
{
cvReleaseImage(&g_ForeImg);
}
if (g_SegementMask)
{
cvReleaseImage(&g_SegementMask);
}
if (g_UpdateMask)
{
cvReleaseImage(&g_UpdateMask);
}
if (g_Gray)
{
cvReleaseImage(&g_Gray);
}
if (g_Model!=NULL)
{
delete[]g_Model;
g_Model=NULL;
}
}
最后附上调用的main函数;
int _tmain(int argc, _TCHAR* argv[])
{
CvCapture *capture=NULL;
IplImage* frame=NULL;
IplImage* pForeImg=NULL;
IplImage* segImg=NULL;
char *file_path="E:\\testVideo\\VTS_01_4.avi"; // m1 test2 锦带河 VTS_01_4_2 head rear VTS_01_6_2 VTS_01_4
//const char* file_path="E:\\suntektechvideo\\锦带河.avi"; //test2
capture=cvCreateFileCapture(file_path);
if (!capture)
{
//cout<<"Read Video File Error!"<<endl;
return -1;
}
frame=cvQueryFrame(capture);
frame=cvQueryFrame(capture);
cvNamedWindow("img",1);
cvNamedWindow("foreN",1);
//cvNamedWindow("seg",1);
Vibe* pV=new Vibe(frame);
while(frame=cvQueryFrame(capture))
{
pV->Detect(frame);
pForeImg=pV->GetForeground();
//segImg=pV->GetSegMask();
//frame->origin=1;
//pForeImg->origin=1;
cvShowImage("img",frame);
cvShowImage("foreN",pForeImg);
//cvShowImage("seg",segImg);
cvWaitKey(1);
}
cvReleaseImage(&frame);
cvReleaseImage(&pForeImg);
cvReleaseCapture(&capture);
return 0;
}
代码没做过多的注释,但现有的注释应该对于理解代码足够了。另外,对于计算机视觉里的任何一种算法都不是万能的,VIBE也不例外,只能说VIBE相对其他算法有一定的优势,但是还是有相当的不足,其pixel-wise-based的灰度建模方式解决不了pixel-wise建模算法共有的问题,其他必要辅助信息的融合是必要的。