paper 83：前景检测算法_1(codebook和平均背景法)
前景分割中一个非常重要的研究方向就是背景减图法，因为背景减图的方法简单，原理容易被想到，且在智能视频监控领域中，摄像机很多情况下是固定的，且背景也是基本不变或者是缓慢变换的，在这种场合背景减图法的应用驱使了其不少科研人员去研究它。
但是背景减图获得前景图像的方法缺点也很多：比如说光照因素，遮挡因素，动态周期背景，且背景非周期背景，且一般情况下我们考虑的是每个像素点之间独立，这对实际应用留下了很大的隐患。
这一小讲主要是讲简单背景减图法和codebook法。
一、简单背景减图法的工作原理。
在视频对背景进行建模的过程中，每2帧图像之间对应像素点灰度值算出一个误差值，在背景建模时间内算出该像素点的平均值，误差平均值，然后在平均差值的基础上+-误差平均值的常数(这个系数需要手动调整)倍作为背景图像的阈值范围，所以当进行前景检测时，当相应点位置来了一个像素时，如果来的这个像素的每个通道的灰度值都在这个阈值范围内，则认为是背景用0表示，否则认为是前景用255表示。
下面的一个工程是learning opencv一书中作者提供的源代码，关于简单背景减图的代码和注释如下：
avg_background.h文件：
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////// // Accumulate average and ~std (really absolute difference) image and use this to detect background and foreground // // Typical way of using this is to: // AllocateImages(); ////loop for N images to accumulate background differences // accumulateBackground(); ////When done, turn this into our avg and std model with high and low bounds // createModelsfromStats(); ////Then use the function to return background in a mask (255 == foreground, 0 == background) // backgroundDiff(IplImage *I,IplImage *Imask, int num); ////Then tune the high and low difference from average image background acceptance thresholds // float scalehigh,scalelow; //Set these, defaults are 7 and 6. Note: scalelow is how many average differences below average // scaleHigh(scalehigh); // scaleLow(scalelow); ////That is, change the scale high and low bounds for what should be background to make it work. ////Then continue detecting foreground in the mask image // backgroundDiff(IplImage *I,IplImage *Imask, int num); // //NOTES: num is camera number which varies from 0 ... NUM_CAMERAS - 1. Typically you only have one camera, but this routine allows // you to index many. // #ifndef AVGSEG_ #define AVGSEG_ #include "cv.h" // define all of the opencv classes etc. #include "highgui.h" #include "cxcore.h" //IMPORTANT DEFINES: #define NUM_CAMERAS 1 //This function can handle an array of cameras #define HIGH_SCALE_NUM 7.0 //How many average differences from average image on the high side == background #define LOW_SCALE_NUM 6.0 //How many average differences from average image on the low side == background void AllocateImages(IplImage *I); void DeallocateImages(); void accumulateBackground(IplImage *I, int number=0); void scaleHigh(float scale = HIGH_SCALE_NUM, int num = 0); void scaleLow(float scale = LOW_SCALE_NUM, int num = 0); void createModelsfromStats(); void backgroundDiff(IplImage *I,IplImage *Imask, int num = 0); #endif
avg_background.cpp文件:
// avg_background.cpp : 定义控制台应用程序的入口点。
//
#include "stdafx.h"
#include "avg_background.h"
//GLOBALS
IplImage *IavgF[NUM_CAMERAS],*IdiffF[NUM_CAMERAS], *IprevF[NUM_CAMERAS], *IhiF[NUM_CAMERAS], *IlowF[NUM_CAMERAS];
IplImage *Iscratch,*Iscratch2,*Igray1,*Igray2,*Igray3,*Imaskt;
IplImage *Ilow1[NUM_CAMERAS],*Ilow2[NUM_CAMERAS],*Ilow3[NUM_CAMERAS],*Ihi1[NUM_CAMERAS],*Ihi2[NUM_CAMERAS],*Ihi3[NUM_CAMERAS];
float Icount[NUM_CAMERAS];
void AllocateImages(IplImage *I) //I is just a sample for allocation purposes
{
for(int i = 0; i<NUM_CAMERAS; i++){
IavgF[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 3 );
IdiffF[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 3 );
IprevF[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 3 );
IhiF[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 3 );
IlowF[i] = cvCreateImage(cvGetSize(I), IPL_DEPTH_32F, 3 );
Ilow1[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Ilow2[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Ilow3[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Ihi1[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Ihi2[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Ihi3[i] = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
cvZero(IavgF[i] );
cvZero(IdiffF[i] );
cvZero(IprevF[i] );
cvZero(IhiF[i] );
cvZero(IlowF[i] );
Icount[i] = 0.00001; //Protect against divide by zero
}
Iscratch = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 3 );
Iscratch2 = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 3 );
Igray1 = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Igray2 = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Igray3 = cvCreateImage( cvGetSize(I), IPL_DEPTH_32F, 1 );
Imaskt = cvCreateImage( cvGetSize(I), IPL_DEPTH_8U, 1 );
cvZero(Iscratch);
cvZero(Iscratch2 );
}
void DeallocateImages()
{
for(int i=0; i<NUM_CAMERAS; i++){
cvReleaseImage(&IavgF[i]);
cvReleaseImage(&IdiffF[i] );
cvReleaseImage(&IprevF[i] );
cvReleaseImage(&IhiF[i] );
cvReleaseImage(&IlowF[i] );
cvReleaseImage(&Ilow1[i] );
cvReleaseImage(&Ilow2[i] );
cvReleaseImage(&Ilow3[i] );
cvReleaseImage(&Ihi1[i] );
cvReleaseImage(&Ihi2[i] );
cvReleaseImage(&Ihi3[i] );
}
cvReleaseImage(&Iscratch);
cvReleaseImage(&Iscratch2);
cvReleaseImage(&Igray1 );
cvReleaseImage(&Igray2 );
cvReleaseImage(&Igray3 );
cvReleaseImage(&Imaskt);
}
// Accumulate the background statistics for one more frame
// We accumulate the images, the image differences and the count of images for the
// the routine createModelsfromStats() to work on after we're done accumulating N frames.
// I Background image, 3 channel, 8u
// number Camera number
void accumulateBackground(IplImage *I, int number)
{
static int first = 1;
cvCvtScale(I,Iscratch,1,0); //To float;#define cvCvtScale cvConvertScale #define cvScale cvConvertScale
if (!first){
cvAcc(Iscratch,IavgF[number]);//将2幅图像相加：IavgF[number]=IavgF[number]+Iscratch，IavgF[]里面装的是时间序列图片的累加
cvAbsDiff(Iscratch,IprevF[number],Iscratch2);//将2幅图像相减：Iscratch2=abs(Iscratch-IprevF[number]);
cvAcc(Iscratch2,IdiffF[number]);//IdiffF[]里面装的是图像差的累积和
Icount[number] += 1.0;//累积的图片帧数计数
}
first = 0;
cvCopy(Iscratch,IprevF[number]);//执行完该函数后，将当前帧数据保存为前一帧数据
}
// Scale the average difference from the average image high acceptance threshold
void scaleHigh(float scale, int num)//设定背景建模时的高阈值函数
{
cvConvertScale(IdiffF[num],Iscratch,scale); //Converts with rounding and saturation
cvAdd(Iscratch,IavgF[num],IhiF[num]);//将平均累积图像与误差累积图像缩放scale倍然后再相加
cvCvtPixToPlane( IhiF[num], Ihi1[num],Ihi2[num],Ihi3[num], 0 );//#define cvCvtPixToPlane cvSplit,且cvSplit是将一个多通道矩阵转换为几个单通道矩阵
}
// Scale the average difference from the average image low acceptance threshold
void scaleLow(float scale, int num)//设定背景建模时的低阈值函数
{
cvConvertScale(IdiffF[num],Iscratch,scale); //Converts with rounding and saturation
cvSub(IavgF[num],Iscratch,IlowF[num]);//将平均累积图像与误差累积图像缩放scale倍然后再相减
cvCvtPixToPlane( IlowF[num], Ilow1[num],Ilow2[num],Ilow3[num], 0 );
}
//Once you've learned the background long enough, turn it into a background model
void createModelsfromStats()
{
for(int i=0; i<NUM_CAMERAS; i++)
{
cvConvertScale(IavgF[i],IavgF[i],(double)(1.0/Icount[i]));//此处为求出累积求和图像的平均值
cvConvertScale(IdiffF[i],IdiffF[i],(double)(1.0/Icount[i]));//此处为求出累计误差图像的平均值
cvAddS(IdiffF[i],cvScalar(1.0,1.0,1.0),IdiffF[i]); //Make sure diff is always something，cvAddS是用于一个数值和一个标量相加
scaleHigh(HIGH_SCALE_NUM,i);//HIGH_SCALE_NUM初始定义为7，其实就是一个倍数
scaleLow(LOW_SCALE_NUM,i);//LOW_SCALE_NUM初始定义为6
}
}
// Create a binary: 0,255 mask where 255 means forground pixel
// I Input image, 3 channel, 8u
// Imask mask image to be created, 1 channel 8u
// num camera number.
//
void backgroundDiff(IplImage *I,IplImage *Imask, int num) //Mask should be grayscale
{
cvCvtScale(I,Iscratch,1,0); //To float;
//Channel 1
cvCvtPixToPlane( Iscratch, Igray1,Igray2,Igray3, 0 );
cvInRange(Igray1,Ilow1[num],Ihi1[num],Imask);//Igray1[]中相应的点在Ilow1[]和Ihi1[]之间时，Imask中相应的点为255(背景符合)
//Channel 2
cvInRange(Igray2,Ilow2[num],Ihi2[num],Imaskt);//也就是说对于每一幅图像的绝对值差小于绝对值差平均值的6倍或者大于绝对值差平均值的7倍被认为是前景图像
cvOr(Imask,Imaskt,Imask);
//Channel 3
cvInRange(Igray3,Ilow3[num],Ihi3[num],Imaskt);//这里的固定阈值6和7太不合理了，还好工程后面可以根据实际情况手动调整！
cvOr(Imask,Imaskt,Imask);
//Finally, invert the results
cvSubRS( Imask, cvScalar(255), Imask);//前景用255表示了，背景是用0表示
}
二、codebook算法工作原理
考虑到简单背景减图法无法对动态的背景建模，有学者就提出了codebook算法。
该算法为图像中每一个像素点建立一个码本，每个码本可以包括多个码元，每个码元有它的学习时最大最小阈值，检测时的最大最小阈值等成员。在背景建模期间，每当来了一幅新图片，对每个像素点进行码本匹配，也就是说如果该像素值在码本中某个码元的学习阈值内，则认为它离过去该对应点出现过的历史情况偏离不大，通过一定的像素值比较，如果满足条件，此时还可以更新对应点的学习阈值和检测阈值。如果新来的像素值对码本中每个码元都不匹配，则有可能是由于背景是动态的，所以我们需要为其建立一个新的码元，并且设置相应的码元成员变量。因此，在背景学习的过程中，每个像素点可以对应多个码元，这样就可以学到复杂的动态背景。
关于codebook算法的代码和注释如下：
cv_yuv_codebook.h文件：
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////// // Accumulate average and ~std (really absolute difference) image and use this to detect background and foreground // // Typical way of using this is to: // AllocateImages(); ////loop for N images to accumulate background differences // accumulateBackground(); ////When done, turn this into our avg and std model with high and low bounds // createModelsfromStats(); ////Then use the function to return background in a mask (255 == foreground, 0 == background) // backgroundDiff(IplImage *I,IplImage *Imask, int num); ////Then tune the high and low difference from average image background acceptance thresholds // float scalehigh,scalelow; //Set these, defaults are 7 and 6. Note: scalelow is how many average differences below average // scaleHigh(scalehigh); // scaleLow(scalelow); ////That is, change the scale high and low bounds for what should be background to make it work. ////Then continue detecting foreground in the mask image // backgroundDiff(IplImage *I,IplImage *Imask, int num); // //NOTES: num is camera number which varies from 0 ... NUM_CAMERAS - 1. Typically you only have one camera, but this routine allows // you to index many. // #ifndef AVGSEG_ #define AVGSEG_ #include "cv.h" // define all of the opencv classes etc. #include "highgui.h" #include "cxcore.h" //IMPORTANT DEFINES: #define NUM_CAMERAS 1 //This function can handle an array of cameras #define HIGH_SCALE_NUM 7.0 //How many average differences from average image on the high side == background #define LOW_SCALE_NUM 6.0 //How many average differences from average image on the low side == background void AllocateImages(IplImage *I); void DeallocateImages(); void accumulateBackground(IplImage *I, int number=0); void scaleHigh(float scale = HIGH_SCALE_NUM, int num = 0); void scaleLow(float scale = LOW_SCALE_NUM, int num = 0); void createModelsfromStats(); void backgroundDiff(IplImage *I,IplImage *Imask, int num = 0); #endif
cv_yuv_codebook.cpp文件：
////////YUV CODEBOOK
// Gary Bradski, July 14, 2005
#include "stdafx.h"
#include "cv_yuv_codebook.h"
//GLOBALS FOR ALL CAMERA MODELS
//For connected components:
int CVCONTOUR_APPROX_LEVEL = 2; // Approx.threshold - the bigger it is, the simpler is the boundary
int CVCLOSE_ITR = 1; // How many iterations of erosion and/or dialation there should be
//#define CVPERIMSCALE 4 // image (width+height)/PERIMSCALE. If contour lenght < this, delete that contour
//For learning background
//Just some convienience macros
#define CV_CVX_WHITE CV_RGB(0xff,0xff,0xff)
#define CV_CVX_BLACK CV_RGB(0x00,0x00,0x00)
///////////////////////////////////////////////////////////////////////////////////
// int updateCodeBook(uchar *p, codeBook &c, unsigned cbBounds)
// Updates the codebook entry with a new data point
//
// p Pointer to a YUV pixel
// c Codebook for this pixel
// cbBounds Learning bounds for codebook (Rule of thumb: 10)
// numChannels Number of color channels we're learning
//
// NOTES:
// cvBounds must be of size cvBounds[numChannels]
//
// RETURN
// codebook index
int cvupdateCodeBook(uchar *p, codeBook &c, unsigned *cbBounds, int numChannels)
{
if(c.numEntries == 0) c.t = 0;//说明每个像素如果遍历了的话至少对应一个码元
c.t += 1; //Record learning event，遍历该像素点的次数加1
//SET HIGH AND LOW BOUNDS
int n;
unsigned int high[3],low[3];
for(n=0; n<numChannels; n++)//为该像素点的每个通道设置最大阈值和最小阈值，后面用来更新学习的高低阈值时有用
{
high[n] = *(p+n)+*(cbBounds+n);
if(high[n] > 255) high[n] = 255;
low[n] = *(p+n)-*(cbBounds+n);
if(low[n] < 0) low[n] = 0;
}
int matchChannel;
//SEE IF THIS FITS AN EXISTING CODEWORD
int i;
for(i=0; i<c.numEntries; i++)//需要对所有的码元进行扫描
{
matchChannel = 0;
for(n=0; n<numChannels; n++)
{
//这个地方要非常小心，if条件不是下面表达的
//if((c.cb[i]->min[n]-c.cb[i]->learnLow[n] <= *(p+n)) && (*(p+n) <= c.cb[i]->max[n]+c.cb[i]->learnHigh[n]))
//原因是因为在每次建立一个新码元的时候，learnHigh[n]和learnLow[n]的范围就在max[n]和min[n]上扩展了cbBounds[n]，所以说
//learnHigh[n]和learnLow[n]的变化范围实际上比max[n]和min[n]的大
if((c.cb[i]->learnLow[n] <= *(p+n)) && (*(p+n) <= c.cb[i]->learnHigh[n])) //Found an entry for this channel
{
matchChannel++;
}
}
if(matchChannel == numChannels) //If an entry was found over all channels，找到了该元素此刻对应的码元
{
c.cb[i]->t_last_update = c.t;
//adjust this codeword for the first channel
//更新每个码元的最大最小阈值，因为这2个阈值在后面的前景分离过程要用到
for(n=0; n<numChannels; n++)
{
if(c.cb[i]->max[n] < *(p+n))//用该点的像素值更新该码元的最大值，所以max[n]保存的是实际上历史出现过的最大像素值
{
c.cb[i]->max[n] = *(p+n);//因为这个for语句是在匹配成功了的条件阈值下的，所以一般来说改变后的max[n]和min[n]
//也不会过学习的高低阈值，并且学习的高低阈值也一直在缓慢变化
}
else if(c.cb[i]->min[n] > *(p+n))//用该点的像素值更新该码元的最小值，所以min[n]保存的是实际上历史出现过的最小像素值
{
c.cb[i]->min[n] = *(p+n);
}
}
break;//一旦找到了该像素的一个码元后就不用继续往后找了，加快算法速度。因为最多只有一个码元与之对应
}
}
//OVERHEAD TO TRACK POTENTIAL STALE ENTRIES
for(int s=0; s<c.numEntries; s++)
{
//This garbage is to track which codebook entries are going stale
int negRun = c.t - c.cb[s]->t_last_update;//negRun表示码元没有更新的时间间隔
if(c.cb[s]->stale < negRun) c.cb[s]->stale = negRun;//更新每个码元的statle
}
//ENTER A NEW CODE WORD IF NEEDED
if(i == c.numEntries) //No existing code word found, make a new one，只有当该像素码本中的所有码元都不符合要求时才满足if条件
{
code_element **foo = new code_element* [c.numEntries+1];//创建一个新的码元序列
for(int ii=0; ii<c.numEntries; ii++)
{
foo[ii] = c.cb[ii];//将码本前面所有的码元地址赋给foo
}
foo[c.numEntries] = new code_element;//创建一个新码元并赋给foo指针的下一个空位
if(c.numEntries) delete [] c.cb;//？
c.cb = foo;
for(n=0; n<numChannels; n++)//给新建立的码元结构体元素赋值
{
c.cb[c.numEntries]->learnHigh[n] = high[n];//当建立一个新码元时，用当前值附近cbBounds范围作为码元box的学习阈值
c.cb[c.numEntries]->learnLow[n] = low[n];
c.cb[c.numEntries]->max[n] = *(p+n);//当建立一个新码元时，用当前值作为码元box的最大最小边界值
c.cb[c.numEntries]->min[n] = *(p+n);
}
c.cb[c.numEntries]->t_last_update = c.t;
c.cb[c.numEntries]->stale = 0;//因为刚建立，所有为0
c.numEntries += 1;//码元的个数加1
}
//SLOWLY ADJUST LEARNING BOUNDS
for(n=0; n<numChannels; n++)//每次遍历该像素点就将每个码元的学习最大阈值变大，最小阈值变小，但是都是缓慢变化的
{ //如果是新建立的码元，则if条件肯定不满足
if(c.cb[i]->learnHigh[n] < high[n]) c.cb[i]->learnHigh[n] += 1;
if(c.cb[i]->learnLow[n] > low[n]) c.cb[i]->learnLow[n] -= 1;
}
return(i);//返回所找到码本中码元的索引
}
///////////////////////////////////////////////////////////////////////////////////
// uchar cvbackgroundDiff(uchar *p, codeBook &c, int minMod, int maxMod)
// Given a pixel and a code book, determine if the pixel is covered by the codebook
//
// p pixel pointer (YUV interleaved)
// c codebook reference
// numChannels Number of channels we are testing
// maxMod Add this (possibly negative) number onto max level when code_element determining if new pixel is foreground
// minMod Subract this (possible negative) number from min level code_element when determining if pixel is foreground
//
// NOTES:
// minMod and maxMod must have length numChannels, e.g. 3 channels => minMod[3], maxMod[3].
//
// Return
// 0 => background, 255 => foreground
uchar cvbackgroundDiff(uchar *p, codeBook &c, int numChannels, int *minMod, int *maxMod)
{
int matchChannel;
//SEE IF THIS FITS AN EXISTING CODEWORD
int i;
for(i=0; i<c.numEntries; i++)
{
matchChannel = 0;
for(int n=0; n<numChannels; n++)
{
if((c.cb[i]->min[n] - minMod[n] <= *(p+n)) && (*(p+n) <= c.cb[i]->max[n] + maxMod[n]))
{
matchChannel++; //Found an entry for this channel
}
else
{
break;//加快速度，当一个通道不满足时提前结束
}
}
if(matchChannel == numChannels)
{
break; //Found an entry that matched all channels，加快速度，当一个码元找到时，提前结束
}
}
if(i >= c.numEntries) return(255);//255代表前景，因为所有的码元都不满足条件
return(0);//0代表背景，因为至少有一个码元满足条件
}
//UTILITES/////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////
//int clearStaleEntries(codeBook &c)
// After you've learned for some period of time, periodically call this to clear out stale codebook entries
//
//c Codebook to clean up
//
// Return
// number of entries cleared
int cvclearStaleEntries(codeBook &c)//对每一个码本进行检查
{
int staleThresh = c.t>>1;//阈值设置为访问该码元的次数的一半，经验值
int *keep = new int [c.numEntries];
int keepCnt = 0;
//SEE WHICH CODEBOOK ENTRIES ARE TOO STALE
for(int i=0; i<c.numEntries; i++)
{
if(c.cb[i]->stale > staleThresh)//当在背景建模期间有一半的时间内，codebook的码元条目没有被访问，则该条目将被删除
keep[i] = 0; //Mark for destruction
else
{
keep[i] = 1; //Mark to keep，为1时，该码本的条目将被保留
keepCnt += 1;//keepCnt记录了要保持的codebook的数目
}
}
//KEEP ONLY THE GOOD
c.t = 0; //Full reset on stale tracking
code_element **foo = new code_element* [keepCnt];//重新建立一个码本的双指针
int k=0;
for(int ii=0; ii<c.numEntries; ii++)
{
if(keep[ii])
{
foo[k] = c.cb[ii];//要保持该码元的话就要把码元结构体复制到fook
foo[k]->stale = 0; //We have to refresh these entries for next clearStale，不被访问的累加器stale重新赋值0
foo[k]->t_last_update = 0;//
k++;
}
}
//CLEAN UP
delete [] keep;
delete [] c.cb;
c.cb = foo;
int numCleared = c.numEntries - keepCnt;//numCleared中保存的是被删除码元的个数
c.numEntries = keepCnt;//最后新的码元数为保存下来码元的个数
return(numCleared);//返回被删除的码元个数
}
/////////////////////////////////////////////////////////////////////////////////
//int countSegmentation(codeBook *c, IplImage *I)
//
//Count how many pixels are detected as foreground
// c Codebook
// I Image (yuv, 24 bits)
// numChannels Number of channels we are testing
// maxMod Add this (possibly negative) number onto max level when code_element determining if new pixel is foreground
// minMod Subract this (possible negative) number from min level code_element when determining if pixel is foreground
//
// NOTES:
// minMod and maxMod must have length numChannels, e.g. 3 channels => minMod[3], maxMod[3].
//
//Return
// Count of fg pixels
//
int cvcountSegmentation(codeBook *c, IplImage *I, int numChannels, int *minMod, int *maxMod)
{
int count = 0,i;
uchar *pColor;
int imageLen = I->width * I->height;
//GET BASELINE NUMBER OF FG PIXELS FOR Iraw
pColor = (uchar *)((I)->imageData);
for(i=0; i<imageLen; i++)
{
if(cvbackgroundDiff(pColor, c[i], numChannels, minMod, maxMod))//对每一个像素点都要检测其是否为前景，如果是的话，计数器count就加1
count++;
pColor += 3;
}
return(count);//返回图像I的前景像素点的个数
}
///////////////////////////////////////////////////////////////////////////////////////////
//void cvconnectedComponents(IplImage *mask, int poly1_hull0, float perimScale, int *num, CvRect *bbs, CvPoint *centers)
// This cleans up the forground segmentation mask derived from calls to cvbackgroundDiff
//
// mask Is a grayscale (8 bit depth) "raw" mask image which will be cleaned up
//
// OPTIONAL PARAMETERS:
// poly1_hull0 If set, approximate connected component by (DEFAULT) polygon, or else convex hull (0)
// perimScale Len = image (width+height)/perimScale. If contour len < this, delete that contour (DEFAULT: 4)
// num Maximum number of rectangles and/or centers to return, on return, will contain number filled (DEFAULT: NULL)
// bbs Pointer to bounding box rectangle vector of length num. (DEFAULT SETTING: NULL)
// centers Pointer to contour centers vectore of length num (DEFULT: NULL)
//
void cvconnectedComponents(IplImage *mask, int poly1_hull0, float perimScale, int *num, CvRect *bbs, CvPoint *centers)
{
static CvMemStorage* mem_storage = NULL;
static CvSeq* contours = NULL;
//CLEAN UP RAW MASK
//开运算作用：平滑轮廓，去掉细节,断开缺口
cvMorphologyEx( mask, mask, NULL, NULL, CV_MOP_OPEN, CVCLOSE_ITR );//对输入mask进行开操作，CVCLOSE_ITR为开操作的次数，输出为mask图像
//闭运算作用：平滑轮廓，连接缺口
cvMorphologyEx( mask, mask, NULL, NULL, CV_MOP_CLOSE, CVCLOSE_ITR );//对输入mask进行闭操作，CVCLOSE_ITR为闭操作的次数，输出为mask图像
//FIND CONTOURS AROUND ONLY BIGGER REGIONS
if( mem_storage==NULL ) mem_storage = cvCreateMemStorage(0);
else cvClearMemStorage(mem_storage);
//CV_RETR_EXTERNAL=0是在types_c.h中定义的，CV_CHAIN_APPROX_SIMPLE=2也是在该文件中定义的
CvContourScanner scanner = cvStartFindContours(mask,mem_storage,sizeof(CvContour),CV_RETR_EXTERNAL,CV_CHAIN_APPROX_SIMPLE);
CvSeq* c;
int numCont = 0;
while( (c = cvFindNextContour( scanner )) != NULL )
{
double len = cvContourPerimeter( c );
double q = (mask->height + mask->width) /perimScale; //calculate perimeter len threshold
if( len < q ) //Get rid of blob if it's perimeter is too small
{
cvSubstituteContour( scanner, NULL );
}
else //Smooth it's edges if it's large enough
{
CvSeq* c_new;
if(poly1_hull0) //Polygonal approximation of the segmentation
c_new = cvApproxPoly(c,sizeof(CvContour),mem_storage,CV_POLY_APPROX_DP, CVCONTOUR_APPROX_LEVEL,0);
else //Convex Hull of the segmentation
c_new = cvConvexHull2(c,mem_storage,CV_CLOCKWISE,1);
cvSubstituteContour( scanner, c_new );
numCont++;
}
}
contours = cvEndFindContours( &scanner );
// PAINT THE FOUND REGIONS BACK INTO THE IMAGE
cvZero( mask );
IplImage *maskTemp;
//CALC CENTER OF MASS AND OR BOUNDING RECTANGLES
if(num != NULL)
{
int N = *num, numFilled = 0, i=0;
CvMoments moments;
double M00, M01, M10;
maskTemp = cvCloneImage(mask);
for(i=0, c=contours; c != NULL; c = c->h_next,i++ )
{
if(i < N) //Only process up to *num of them
{
cvDrawContours(maskTemp,c,CV_CVX_WHITE, CV_CVX_WHITE,-1,CV_FILLED,8);
//Find the center of each contour
if(centers != NULL)
{
cvMoments(maskTemp,&moments,1);
M00 = cvGetSpatialMoment(&moments,0,0);
M10 = cvGetSpatialMoment(&moments,1,0);
M01 = cvGetSpatialMoment(&moments,0,1);
centers[i].x = (int)(M10/M00);
centers[i].y = (int)(M01/M00);
}
//Bounding rectangles around blobs
if(bbs != NULL)
{
bbs[i] = cvBoundingRect(c);
}
cvZero(maskTemp);
numFilled++;
}
//Draw filled contours into mask
cvDrawContours(mask,c,CV_CVX_WHITE,CV_CVX_WHITE,-1,CV_FILLED,8); //draw to central mask
} //end looping over contours
*num = numFilled;
cvReleaseImage( &maskTemp);
}
//ELSE JUST DRAW PROCESSED CONTOURS INTO THE MASK
else
{
for( c=contours; c != NULL; c = c->h_next )
{
cvDrawContours(mask,c,CV_CVX_WHITE, CV_CVX_BLACK,-1,CV_FILLED,8);
}
}
}
三、2种算法进行对比。
Learning Opencv的作者将这两种算法做了下对比，用的视频是有风吹动树枝的动态背景，一段时间过后的前景是视频中移动的手。
当然在这个工程中，作者除了体现上述简单背景差法和codobook算法的一些原理外，还引入了很多细节来优化前景分割效果。比如说误差计算时的方差和协方差计算加速方法，消除像素点内长时间没有被访问过的码元，对检测到的粗糙原始前景图用连通域分析法清楚噪声，其中引入了形态学中的几种操作，使用多边形拟合前景轮廓等细节处理。
在看作者代码前，最好先看下下面几个变量的物理含义。
maxMod[n]：用训练好的背景模型进行前景检测时用到，判断点是否小于max[n] + maxMod[n])。
minMod[n]：用训练好的背景模型进行前景检测时用到，判断点是否小于min[n] -minMod[n])。
cbBounds*：训练背景模型时用到，可以手动输入该参数，这个数主要是配合high[n]和low[n]来用的。
learnHigh[n]：背景学习过程中当一个新像素来时用来判断是否在已有的码元中，是阈值的上界部分。
learnLow[n]：背景学习过程中当一个新像素来时用来判断是否在已有的码元中，是阈值的下界部分。
max[n]： 背景学习过程中每个码元学习到的最大值，在前景分割时配合maxMod[n]用的。
min[n]： 背景学习过程中每个码元学习到的最小值，在前景分割时配合minMod[n]用的。
high[n]：背景学习过程中用来调整learnHigh[n]的，如果learnHigh[n]<high[n],则learnHigh[n]缓慢加1
low[n]： 背景学习过程中用来调整learnLow[n]的，如果learnLow[n]>Low[n],则learnLow[缓慢减1
该工程带主函数部分代码和注释如下：
#include "stdafx.h"
#include "cv.h"
#include "highgui.h"
#include <stdio.h>
#include <stdlib.h>
#include <ctype.h>
#include "avg_background.h"
#include "cv_yuv_codebook.h"
//VARIABLES for CODEBOOK METHOD:
codeBook *cB; //This will be our linear model of the image, a vector
//of lengh = height*width
int maxMod[CHANNELS]; //Add these (possibly negative) number onto max
// level when code_element determining if new pixel is foreground
int minMod[CHANNELS]; //Subract these (possible negative) number from min
//level code_element when determining if pixel is foreground
unsigned cbBounds[CHANNELS]; //Code Book bounds for learning
bool ch[CHANNELS]; //This sets what channels should be adjusted for background bounds
int nChannels = CHANNELS;
int imageLen = 0;
uchar *pColor; //YUV pointer
void help() {
printf("\nLearn background and find foreground using simple average and average difference learning method:\n"
"\nUSAGE:\n ch9_background startFrameCollection# endFrameCollection# [movie filename, else from camera]\n"
"If from AVI, then optionally add HighAvg, LowAvg, HighCB_Y LowCB_Y HighCB_U LowCB_U HighCB_V LowCB_V\n\n"
"***Keep the focus on the video windows, NOT the consol***\n\n"
"INTERACTIVE PARAMETERS:\n"
"\tESC,q,Q - quit the program\n"
"\th - print this help\n"
"\tp - pause toggle\n"
"\ts - single step\n"
"\tr - run mode (single step off)\n"
"=== AVG PARAMS ===\n"
"\t- - bump high threshold UP by 0.25\n"
"\t= - bump high threshold DOWN by 0.25\n"
"\t[ - bump low threshold UP by 0.25\n"
"\t] - bump low threshold DOWN by 0.25\n"
"=== CODEBOOK PARAMS ===\n"
"\ty,u,v- only adjust channel 0(y) or 1(u) or 2(v) respectively\n"
"\ta - adjust all 3 channels at once\n"
"\tb - adjust both 2 and 3 at once\n"
"\ti,o - bump upper threshold up,down by 1\n"
"\tk,l - bump lower threshold up,down by 1\n"
);
}
//
//USAGE: ch9_background startFrameCollection# endFrameCollection# [movie filename, else from camera]
//If from AVI, then optionally add HighAvg, LowAvg, HighCB_Y LowCB_Y HighCB_U LowCB_U HighCB_V LowCB_V
//
int main(int argc, char** argv)
{
IplImage* rawImage = 0, *yuvImage = 0; //yuvImage is for codebook method
IplImage *ImaskAVG = 0,*ImaskAVGCC = 0;
IplImage *ImaskCodeBook = 0,*ImaskCodeBookCC = 0;
CvCapture* capture = 0;
int startcapture = 1;
int endcapture = 30;
int c,n;
maxMod[0] = 3; //Set color thresholds to default values
minMod[0] = 10;
maxMod[1] = 1;
minMod[1] = 1;
maxMod[2] = 1;
minMod[2] = 1;
float scalehigh = HIGH_SCALE_NUM;//默认值为6
float scalelow = LOW_SCALE_NUM;//默认值为7
if(argc < 3) {//只有1个参数或者没有参数时，输出错误，并提示help信息，因为该程序本身就算进去了一个参数
printf("ERROR: Too few parameters\n");
help();
}else{//至少有2个参数才算正确
if(argc == 3){//输入为2个参数的情形是从摄像头输入数据
printf("Capture from Camera\n");
capture = cvCaptureFromCAM( 0 );
}
else {//输入大于2个参数时是从文件中读入视频数据
printf("Capture from file %s\n",argv[3]);//第三个参数是读入视频文件的文件名
// capture = cvCaptureFromFile( argv[3] );
capture = cvCreateFileCapture( argv[3] );
if(!capture) { printf("Couldn't open %s\n",argv[3]); return -1;}//读入视频文件失败
}
if(isdigit(argv[1][0])) { //Start from of background capture
startcapture = atoi(argv[1]);//第一个参数表示视频开始的背景训练时的帧，默认是1
printf("startcapture = %d\n",startcapture);
}
if(isdigit(argv[2][0])) { //End frame of background capture
endcapture = atoi(argv[2]);//第二个参数表示的结束背景训练时的，默认为30
printf("endcapture = %d\n");
}
if(argc > 4){ //See if parameters are set from command line，输入多于4个参数表示后面的算法中用到的参数在这里直接输入
//FOR AVG MODEL
if(argc >= 5){
if(isdigit(argv[4][0])){
scalehigh = (float)atoi(argv[4]);
}
}
if(argc >= 6){
if(isdigit(argv[5][0])){
scalelow = (float)atoi(argv[5]);
}
}
//FOR CODEBOOK MODEL, CHANNEL 0
if(argc >= 7){
if(isdigit(argv[6][0])){
maxMod[0] = atoi(argv[6]);
}
}
if(argc >= 8){
if(isdigit(argv[7][0])){
minMod[0] = atoi(argv[7]);
}
}
//Channel 1
if(argc >= 9){
if(isdigit(argv[8][0])){
maxMod[1] = atoi(argv[8]);
}
}
if(argc >= 10){
if(isdigit(argv[9][0])){
minMod[1] = atoi(argv[9]);
}
}
//Channel 2
if(argc >= 11){
if(isdigit(argv[10][0])){
maxMod[2] = atoi(argv[10]);
}
}
if(argc >= 12){
if(isdigit(argv[11][0])){
minMod[2] = atoi(argv[11]);
}
}
}
}
//MAIN PROCESSING LOOP:
bool pause = false;
bool singlestep = false;
if( capture )
{
cvNamedWindow( "Raw", 1 );//原始视频图像
cvNamedWindow( "AVG_ConnectComp",1);//平均法连通区域分析后的图像
cvNamedWindow( "ForegroundCodeBook",1);//codebook法后图像
cvNamedWindow( "CodeBook_ConnectComp",1);//codebook法连通区域分析后的图像
cvNamedWindow( "ForegroundAVG",1);//平均法后图像
int i = -1;
for(;;)
{
if(!pause){
// if( !cvGrabFrame( capture ))
// break;
// rawImage = cvRetrieveFrame( capture );
rawImage = cvQueryFrame( capture );
++i;//count it
// printf("%d\n",i);
if(!rawImage)
break;
//REMOVE THIS FOR GENERAL OPERATION, JUST A CONVIENIENCE WHEN RUNNING WITH THE SMALL tree.avi file
if(i == 56){//程序开始运行几十帧后自动暂停，以便后面好手动调整参数
pause = 1;
printf("\n\nVideo paused for your convienience at frame 50 to work with demo\n"
"You may adjust parameters, single step or continue running\n\n");
help();
}
}
if(singlestep){
pause = true;
}
//First time:
if(0 == i) {
printf("\n . . . wait for it . . .\n"); //Just in case you wonder why the image is white at first
//AVG METHOD ALLOCATION
AllocateImages(rawImage);//为算法的使用分配内存
scaleHigh(scalehigh);//设定背景建模时的高阈值函数
scaleLow(scalelow);//设定背景建模时的低阈值函数
ImaskAVG = cvCreateImage( cvGetSize(rawImage), IPL_DEPTH_8U, 1 );
ImaskAVGCC = cvCreateImage( cvGetSize(rawImage), IPL_DEPTH_8U, 1 );
cvSet(ImaskAVG,cvScalar(255));
//CODEBOOK METHOD ALLOCATION:
yuvImage = cvCloneImage(rawImage);
ImaskCodeBook = cvCreateImage( cvGetSize(rawImage), IPL_DEPTH_8U, 1 );//用来装前景背景图的，当然只要一个通道的图像即可
ImaskCodeBookCC = cvCreateImage( cvGetSize(rawImage), IPL_DEPTH_8U, 1 );
cvSet(ImaskCodeBook,cvScalar(255));
imageLen = rawImage->width*rawImage->height;
cB = new codeBook [imageLen];//创建一个码本cB数组，每个像素对应一个码本
for(int f = 0; f<imageLen; f++)
{
cB[f].numEntries = 0;//每个码本的初始码元个数赋值为0
}
for(int nc=0; nc<nChannels;nc++)
{
cbBounds[nc] = 10; //Learning bounds factor，初始值为10
}
ch[0] = true; //Allow threshold setting simultaneously for all channels
ch[1] = true;
ch[2] = true;
}
//If we've got an rawImage and are good to go:
if( rawImage )
{
cvCvtColor( rawImage, yuvImage, CV_BGR2YCrCb );//YUV For codebook method
//This is where we build our background model
if( !pause && i >= startcapture && i < endcapture ){
//LEARNING THE AVERAGE AND AVG DIFF BACKGROUND
accumulateBackground(rawImage);//平均法累加过程
//LEARNING THE CODEBOOK BACKGROUND
pColor = (uchar *)((yuvImage)->imageData);//yuvImage矩阵的首位置
for(int c=0; c<imageLen; c++)
{
cvupdateCodeBook(pColor, cB[c], cbBounds, nChannels);//codebook算法建模过程
pColor += 3;
}
}
//When done, create the background model
if(i == endcapture){
createModelsfromStats();//平均法建模过程
}
//Find the foreground if any
if(i >= endcapture) {//endcapture帧后开始检测前景
//FIND FOREGROUND BY AVG METHOD:
backgroundDiff(rawImage,ImaskAVG);
cvCopy(ImaskAVG,ImaskAVGCC);
cvconnectedComponents(ImaskAVGCC);//平均法中的前景清除
//FIND FOREGROUND BY CODEBOOK METHOD
uchar maskPixelCodeBook;
pColor = (uchar *)((yuvImage)->imageData); //3 channel yuv image
uchar *pMask = (uchar *)((ImaskCodeBook)->imageData); //1 channel image
for(int c=0; c<imageLen; c++)
{
maskPixelCodeBook = cvbackgroundDiff(pColor, cB[c], nChannels, minMod, maxMod);//前景返回255，背景返回0
*pMask++ = maskPixelCodeBook;//将前景检测的结果返回到ImaskCodeBook中
pColor += 3;
}
//This part just to visualize bounding boxes and centers if desired
cvCopy(ImaskCodeBook,ImaskCodeBookCC);
cvconnectedComponents(ImaskCodeBookCC);//codebook算法中的前景清除
}
//Display
cvShowImage( "Raw", rawImage );//除了这张是彩色图外，另外4张都是黑白图
cvShowImage( "AVG_ConnectComp",ImaskAVGCC);
cvShowImage( "ForegroundAVG",ImaskAVG);
cvShowImage( "ForegroundCodeBook",ImaskCodeBook);
cvShowImage( "CodeBook_ConnectComp",ImaskCodeBookCC);
//USER INPUT:
c = cvWaitKey(10)&0xFF;
//End processing on ESC, q or Q
if(c == 27 || c == 'q' | c == 'Q')
break;
//Else check for user input
switch(c)
{
case 'h':
help();
break;
case 'p':
pause ^= 1;
break;
case 's':
singlestep = 1;
pause = false;
break;
case 'r':
pause = false;
singlestep = false;
break;
//AVG BACKROUND PARAMS
case '-'://调整scalehigh的参数，scalehigh的物理意义是误差累加的影响因子，其倒数为缩放倍数，加0.25实际上是减小其影响力
if(i > endcapture){
scalehigh += 0.25;
printf("AVG scalehigh=%f\n",scalehigh);
scaleHigh(scalehigh);
}
break;
case '='://scalehigh减少2.5是增加其影响力
if(i > endcapture){
scalehigh -= 0.25;
printf("AVG scalehigh=%f\n",scalehigh);
scaleHigh(scalehigh);
}
break;
case '[':
if(i > endcapture){//设置设定背景建模时的低阈值函数，同上
scalelow += 0.25;
printf("AVG scalelow=%f\n",scalelow);
scaleLow(scalelow);
}
break;
case ']':
if(i > endcapture){
scalelow -= 0.25;
printf("AVG scalelow=%f\n",scalelow);
scaleLow(scalelow);
}
break;
//CODEBOOK PARAMS
case 'y':
case '0'://激活y通道
ch[0] = 1;
ch[1] = 0;
ch[2] = 0;
printf("CodeBook YUV Channels active: ");
for(n=0; n<nChannels; n++)
printf("%d, ",ch[n]);
printf("\n");
break;
case 'u':
case '1'://激活u通道
ch[0] = 0;
ch[1] = 1;
ch[2] = 0;
printf("CodeBook YUV Channels active: ");
for(n=0; n<nChannels; n++)
printf("%d, ",ch[n]);
printf("\n");
break;
case 'v':
case '2'://激活v通道
ch[0] = 0;
ch[1] = 0;
ch[2] = 1;
printf("CodeBook YUV Channels active: ");
for(n=0; n<nChannels; n++)
printf("%d, ",ch[n]);
printf("\n");
break;
case 'a': //All
case '3'://激活所有通道
ch[0] = 1;
ch[1] = 1;
ch[2] = 1;
printf("CodeBook YUV Channels active: ");
for(n=0; n<nChannels; n++)
printf("%d, ",ch[n]);
printf("\n");
break;
case 'b': //both u and v together
ch[0] = 0;
ch[1] = 1;
ch[2] = 1;
printf("CodeBook YUV Channels active: ");
for(n=0; n<nChannels; n++)
printf("%d, ",ch[n]);
printf("\n");
break;
case 'i': //modify max classification bounds (max bound goes higher)
for(n=0; n<nChannels; n++){//maxMod和minMod是最大值和最小值跳动的阈值
if(ch[n])
maxMod[n] += 1;
printf("%.4d,",maxMod[n]);
}
printf(" CodeBook High Side\n");
break;
case 'o': //modify max classification bounds (max bound goes lower)
for(n=0; n<nChannels; n++){
if(ch[n])
maxMod[n] -= 1;
printf("%.4d,",maxMod[n]);
}
printf(" CodeBook High Side\n");
break;
case 'k': //modify min classification bounds (min bound goes lower)
for(n=0; n<nChannels; n++){
if(ch[n])
minMod[n] += 1;
printf("%.4d,",minMod[n]);
}
printf(" CodeBook Low Side\n");
break;
case 'l': //modify min classification bounds (min bound goes higher)
for(n=0; n<nChannels; n++){
if(ch[n])
minMod[n] -= 1;
printf("%.4d,",minMod[n]);
}
printf(" CodeBook Low Side\n");
break;
}
}
}
cvReleaseCapture( &capture );
cvDestroyWindow( "Raw" );
cvDestroyWindow( "ForegroundAVG" );
cvDestroyWindow( "AVG_ConnectComp");
cvDestroyWindow( "ForegroundCodeBook");
cvDestroyWindow( "CodeBook_ConnectComp");
DeallocateImages();//释放平均法背景建模过程中用到的内存
if(yuvImage) cvReleaseImage(&yuvImage);
if(ImaskAVG) cvReleaseImage(&ImaskAVG);
if(ImaskAVGCC) cvReleaseImage(&ImaskAVGCC);
if(ImaskCodeBook) cvReleaseImage(&ImaskCodeBook);
if(ImaskCodeBookCC) cvReleaseImage(&ImaskCodeBookCC);
delete [] cB;
}
else{ printf("\n\nDarn, Something wrong with the parameters\n\n"); help();
}
return 0;
}
运行结果截图如下：
训练过程视频原图截图：
测试过程视频原图截图：
前景检测过程截图：
可以看到左边2幅截图的对比，codebook算法的效果明显比简单减图法要好，手型比较清晰些。
四、参考文献
Bradski, G. and A. Kaehler (2008). Learning OpenCV: Computer vision with the OpenCV library, O'Reilly Media.
