【背景建模】SACON http://www.cnblogs.com/dwdxdy/p/3530862.html
【背景建模】VIBE http://www.cnblogs.com/dwdxdy/p/3527891.html
创新点：
1).引入控制论的思想，使前景判断阈值和背景模型更新率自适应变化，随背景的复杂程度变化。
2).引入背景复杂程度的度量方法，根据背景复杂程度调整前景判断阈值和背景模型更新率。
基本框架：
R(xi)为自适应的阈值，T(xi)为自适应的更新率，dmin(xi)为背景复杂度的度量。
主要步骤：
1).建立背景模型
PBAS算法采用类似SACON算法背景建模方法，收集前N帧像素以及梯度幅值作为背景模型；
2).前景检测
其检测过程与VIBE算法类似，计算样本集合中满足匹配条件的样本个数，若小于阈值#min，则表示为背景，不同之处在于不同像素的判断阈值R(xi)不同；
B(xi)是背景模型，表示像素点xi的样本集合，F(xi)是前景图像，表示像素点xi的判断结果。
3).背景模型更新
其更新过程与VIBE算法类似，随机选择需要被替换的样本，随机选择像素邻域的样本集合更新，不同之处在于更新率不同，VIBE是固定的更新率，而PBAS的更新率是自适应的，并且更新邻域样本集合是用邻域的新像素值进行更新，而不是用新像素值本身。
4).背景复杂度计算
当新像素值与样本集合的最小距离小于阈值时，该说明该像素可能为背景，此时最小距离可以描述背景的复杂度，距离越大，背景越复杂。因而，采用样本集合中样本更新时的最小距离的平均值作为背景复杂度的度量。
5).自适应调整判断阈值
背景越复杂，其判断阈值应该越大，保证背景像素不被误判为前景，背景越简单，微小的变化都是由前景引起，判断阈值应该越小，调整策略如下：
Rinc/dec是阈值的变化量，Rscale用于控制背景复杂度对判断阈值的调整。
6).自适应调整更新率
背景越复杂，其出现错误判断的概率越大，对应的背景更新应该越少，减少错误判断对背景模型的影响，其更新策略如下：
F(xi)表示前景检测结果，Tinc和Tdec分别表示更新率增加、减小的幅度，Tlower和Tupper分别表示更新率上下限。
算法实现注意事项：
1).参数的设置，可以根据实际的应用调整参数的大小，一般的参数设置如下：
N = 35, #min = 2, Rinc/dec = 0.05, Rlower = 18, Rscale = 5, Tdec = 0.05, Tinc = 1, Tlower = 2, Tupper = 200
2).基于像素的前景判断、背景复杂度计算、判断阈值和更新率的更新计算，计算复杂度，比较耗时，建议采用GPU通用计算技术实现。
参考资料：
Background Segmentation with Feedback: The Pixel-Based Adaptive Segmenter
http://www.mmk.ei.tum.de/~hom/pbas
pbas.h
#include <iostream>
#include "cv.h"
#include "highgui.h"
#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#pragma once
class PBAS
{
public:
PBAS(void);
~PBAS(void);
bool process(cv::Mat* input, cv::Mat* output);
//参数设置
void setN(int);
void setRaute_min(int);
void setR_lower(double);
void setR_incdec(double);
void setR_scale(double);
void setT_init(double);
void setT_lower(double);
void setT_upper(double);
void setT_dec(double);
void setT_inc(double);
void setAlpha(double);
void setBeta(double);
bool isMovement();
private:
void calculateFeatures(std::vector<cv::Mat>* feature, cv::Mat* inputImage);
void checkValid(int *x, int *y);
void decisionThresholdRegulator(float* pt, float* meanDistArr);
void learningRateRegulator(float* pt, float* meanDist, uchar* isFore);
void init(cv::Mat*);
void newInitialization();
//存储D(Xi)的均值
cv::Mat meanMinDist;
float* meanMinDist_Pt;
int width, height;
int chans;
//balance of feature pixel value to conture value 公式5
double alpha, beta;
//##################################################################################
double formerMeanNorm;
//define value of foreground/background pixels
int foregroundValue, backgroundValue;
//##################################################################################
//random number parameters
//random number generator
cv::RNG randomGenerator;
//length of random array initialization
long countOfRandomNumb;
//pre - initialize the randomNumbers for better performance
std::vector<int> randomN, randomMinDist, randomX, randomY, randomT, randomTN;
//###################################################################################
//check if something is moving
bool isMove;
//for init, count number of runs
int runs;
//存储分割模板
cv::Mat* resultMap;
uchar* resultMap_Pt;
//计算当前输入帧特征，包括梯度+灰度值
std::vector<cv::Mat> currentFeatures;
std::vector<float*> currentFeaturesM_Pt;
std::vector<uchar*> currentFeaturesC_Pt;
// background model
std::vector<std::vector<cv::Mat> > backgroundModel;
//背景模型 N 个features
std::vector<std::vector<float*> >B_Mag_Pts;
std::vector<std::vector<uchar*> >B_Col_Pts;
//计算Paper公式5中的Im
double sumMagnitude;
double formerMeanMag;
float formerDistanceBack;
//####################################################################################
//N - Number: Defining the size of the background-history-model
// number of samples per pixel
//size of background history B(x_i)
int N;
//####################################################################################
//####################################################################################
//R-Threshhold - Variables
//minimal Threshold for foreground and initial value of R(x_i)
// radius of the sphere -> lower border boundary
double R_lower;
//factor which defines new threshold of R(x_i) together with meanMinDist(x_i)
// scale for the sphere threshhold to define pixel-based Thresholds
double R_scale;
//decreasing / increasing factor of R(x_i)
// increasing/decreasing factor for the r-Threshold based on the result of rTreshScale * meanMinDistBackground
double R_incdec;
cv::Mat actualR;
float* actualR_Pt; //new pixel-based r-threshhold -> pointer to arrays
//#####################################################################################
//####################################################################################
//counter for minimal distance to background
// Defining the number of background-model-images, which have a lowerDIstance to the current Image than defined by the R-Thresholds, that are necessary
// to decide that this pixel is background
int Raute_min;
//#####################################################################################
//####################################################################################
//initial value of inverse update factor T(x_i)
// Initialize the background-model update rate
double T_init;
//increasing Factor of the update rate 1/T(x_i)
// scale that defines the increasing of the update rate of the background model, if the current pixel is background
//--> more frequently updates if pixel is background because, there shouln't be any change
double T_inc;
//upper boundary of T(x_i)
// defining an upper value, that nrSubsampling can achieve, thus it doesn't reach to an infinite value, where actually no update is possible
// at all
double T_upper;
//lower boundary of T(x_i)
// defining a minimum value for nrSubsampling --> base value 2.0
double T_lower;
//decreasing factor of the update rate 1/T(x_i)
// opposite scale to increasingRateScale, for decreasing the update rate of the background model, if the current pixel is foreground
//--> Thesis: Our Foreground is a real moving object -> thus the background-model is good, so don't update it
double T_dec;
//holds update rate of current pixel
cv::Mat actualT;
float* actualT_Pt;
//#####################################################################################
cv::Mat sobelX, sobelY;
};
pbas.cpp
#include "PBAS.h"
PBAS::PBAS(void) : N(20), R_lower(18), Raute_min(2), T_lower(2), T_upper(200), R_scale(5), R_incdec(0.05), T_dec(0.05), T_inc(1.0)
{
std::cout << "PBAS()" << std::endl;
//feature vector
alpha = 7.0;
beta = 1.0;
formerMeanNorm = 0;
width = 0;
//result image
foregroundValue = 255;
backgroundValue = 0;
//length of random array
countOfRandomNumb = 1000;
//the T(x_i) value needs initiation
T_init = R_lower;
//check if something is moving in the picture
isMove = false;
//for init, count number of runs
runs = 0;
newInitialization();
}
void PBAS::newInitialization()
{
//清空存储各项随机数vector
if (!randomN.empty())
randomN.clear();
if (!randomX.empty())
randomX.clear();
if (!randomY.empty())
randomY.clear();
if (!randomMinDist.empty())
randomMinDist.clear();
if (!randomT.empty())
randomT.clear();
if (!randomTN.empty())
randomTN.clear();
//先分别产生countOfRandomNumb个随机数，提高算法速度
for (int l = 0; l < countOfRandomNumb; l++)
{
randomN.push_back((int)randomGenerator.uniform((int)0, (int)N));
randomX.push_back((int)randomGenerator.uniform(-1, +2));
randomY.push_back((int)randomGenerator.uniform(-1, +2));
randomMinDist.push_back((int)randomGenerator.uniform((int)0, (int)N));
randomT.push_back((int)randomGenerator.uniform((int)0, (int)T_upper));
randomTN.push_back((int)randomGenerator.uniform((int)0, (int)T_upper));
}
}
PBAS::~PBAS(void)
{
std::cout << "~PBAS()" << std::endl;
//以下都可以删除......不明白作者为什么写上去
randomN.clear();
randomX.clear();
randomY.clear();
randomMinDist.clear();
randomT.clear();
randomTN.clear();
for (int k = 0; k < backgroundModel.size(); ++k)
{
if (chans == 1)
{
backgroundModel.at(k).at(0).release();
backgroundModel.at(k).at(1).release();
}
else
{
backgroundModel.at(k).at(0).release();
backgroundModel.at(k).at(1).release();
backgroundModel.at(k).at(2).release();
backgroundModel.at(k).at(3).release();
backgroundModel.at(k).at(4).release();
backgroundModel.at(k).at(5).release();
}
}
backgroundModel.clear();
meanMinDist.release();
actualR.release();
actualT.release();
sobelX.release();
sobelY.release();
}
bool PBAS::process(cv::Mat* input, cv::Mat* output)
{
//第一帧先存储图像rows，cols，chans
if (width != input->cols)
{
width = input->cols;
chans = input->channels();
height = input->rows;
if (input->rows < 1 || input->cols < 1)
{
std::cout << "Error: Occurrence of to small (or empty?) image size in PBAS. STOPPING " << std::endl;
return false;
}
}
//iniate the background model，使用前N帧建立背景模型，次函数调用N次
init(input);
//分割mask
resultMap = new cv::Mat(input->rows, input->cols, CV_8UC1);
//calculate features
//计算当前图像Features，（梯度，灰度）
calculateFeatures(¤tFeatures, input);
//set sumMagnitude to zero at beginning and then sum up in the loop
//计算公式5用到的Im
sumMagnitude = 0;
long glCounterFore = 0;
isMove = false;
//Here starts the whole processing of each pixel of the image
// for each pixel
for (int j = 0; j < resultMap->rows; ++j)
{
//分割mask行指针
resultMap_Pt = resultMap->ptr<uchar>(j);
//对当前特征图(彩色6个Mat，灰度2个Mat)
currentFeaturesM_Pt.clear();
currentFeaturesC_Pt.clear();
std::vector<float*> fT;
std::vector<uchar*> uT;
B_Mag_Pts.clear();
B_Col_Pts.clear();
for (int z = 0; z < chans; ++z)
{
//features vector中的梯度图的行指针
currentFeaturesM_Pt.push_back(currentFeatures.at(z).ptr<float>(j));
//feature vector中的灰度图的行指针
currentFeaturesC_Pt.push_back(currentFeatures.at(z + chans).ptr<uchar>(j));
//背景模型
B_Mag_Pts.push_back(fT);
B_Col_Pts.push_back(uT);
}
//同样取各个行指针
meanMinDist_Pt = meanMinDist.ptr<float>(j);
actualR_Pt = actualR.ptr<float>(j);
actualT_Pt = actualT.ptr<float>(j);
//如果是彩色图像，B_Mag_Pts，B_Col_Pts就分别有3个vector<float*>, 3个vector<uchar*>
//其中每个对应一个通道
//而一个vector<float*>存储runs帧图像的行指针..比如runs=N，则有N个float*，分别对应每个背景特征图
//中梯度图的行地址，同理B_Col_Pts则是灰度图
for (int k = 0; k < runs; ++k)
{
for (int z = 0; z < chans; ++z)
{
B_Mag_Pts.at(z).push_back(backgroundModel.at(k).at(z).ptr<float>(j));
B_Col_Pts.at(z).push_back(backgroundModel.at(k).at(z + chans).ptr<uchar>(j));
}
}
//对每个像素依次操作如下
for (int i = 0; i < resultMap->cols; ++i)
{
//Compare each pixel to in the worst runtime-case each background model
int count = 0;
int index = 0;
double norm = 0.0;
double dist = 0.0;
double minDist = 1000.0;
//在countOfRandomNumb中随机取一个数
int entry = randomGenerator.uniform(3, countOfRandomNumb - 4);
do
{
//如果是彩色图像，
if (chans == 3)
{
//R,G,B通道分别： 第index+1个背景图中的第i+1个位置依次与 当前图像的当前位置，比较梯度
//sqrt(Mr^2 + Mg^2 + Mb^2) 其中Mi表示对应通道的梯度差值
norm = sqrt(
(((double)B_Mag_Pts.at(0).at(index)[i] - ((double)*currentFeaturesM_Pt.at(0))) * ((double)B_Mag_Pts.at(0).at(index)[i] - ((double)*currentFeaturesM_Pt.at(0)))) +
(((double)B_Mag_Pts.at(1).at(index)[i] - ((double)*currentFeaturesM_Pt.at(1))) * ((double)B_Mag_Pts.at(1).at(index)[i] - ((double)*currentFeaturesM_Pt.at(1)))) +
(((double)B_Mag_Pts.at(2).at(index)[i] - ((double)*currentFeaturesM_Pt.at(2))) * ((double)B_Mag_Pts.at(2).at(index)[i] - ((double)*currentFeaturesM_Pt.at(2))))
);
//同上，只不过此处是灰度值的比较
dist = sqrt(
(((double)B_Col_Pts.at(0).at(index)[i] - ((double)*currentFeaturesC_Pt.at(0))) * ((double)B_Col_Pts.at(0).at(index)[i] - ((double)*currentFeaturesC_Pt.at(0)))) +
(((double)B_Col_Pts.at(1).at(index)[i] - ((double)*currentFeaturesC_Pt.at(1))) * ((double)B_Col_Pts.at(1).at(index)[i] - ((double)*currentFeaturesC_Pt.at(1)))) +
(((double)B_Col_Pts.at(2).at(index)[i] - ((double)*currentFeaturesC_Pt.at(2))) * ((double)B_Col_Pts.at(2).at(index)[i] - ((double)*currentFeaturesC_Pt.at(2))))
);
}
else //灰度图像
{
//比上面更简单 abs(Mb - Mc)^2
norm = abs(
(((double)B_Mag_Pts.at(0).at(index)[i] - ((double)*currentFeaturesM_Pt.at(0))) * ((double)B_Mag_Pts.at(0).at(index)[i] - ((double)*currentFeaturesM_Pt.at(0))))
);
dist = abs(
(((double)B_Col_Pts.at(0).at(index)[i] - ((double)*currentFeaturesC_Pt.at(0))) * ((double)B_Col_Pts.at(0).at(index)[i] - ((double)*currentFeaturesC_Pt.at(0))))
);
}
//根据公式5计算距离
dist = ((double)alpha*(norm / formerMeanMag) + beta*dist);
//如果小于距离阈值，记录，同时匹配数+1
if ((dist < *actualR_Pt))
{
++count;
if (minDist > dist)
minDist = dist;
}
//否则，记下当前位置的梯度值，【未解决】
//此处有点奇怪，公式5是说Im表示前一帧图像的梯度平均，为什么这里只计算不匹配背景的位置梯度？？？）
//我试过通过累计所有位置的梯度，然后除以像素个数，好像效果没多大区别...alpha值我没动过，可能调下alpha
else
{
sumMagnitude += (double)(norm);
++glCounterFore;
}
++index;
} while ((count < Raute_min) && (index < runs)); //当前匹配数小于2或者背景没有遍历完，循环以上过程
//#############################################
//update backgroundmodel
// is BACKGROUND 如果是背景
if (count >= Raute_min)
{
*resultMap_Pt = 0;
//使用这种方式替代随机数产生器，提高算法速度
//比如T_upper=200， actualT=2；则ratio=100， 那么ratio小于 randomT的概率多大呢，1/2，对吧。 和咱们的
//更新率1/T = 1/2 一样...其他数值同理
double ratio = std::ceil((double)T_upper / (double)(*actualT_Pt));
//in the first run every distance is zero, because there is no background model
//in the secont run, we have already one image as background model, hence a
// reasonable minDist could be found -> because of the partly 1/run changing in the running average, we set in the first try meanMinDist to the actual minDist value
//存储当前minDist到meanMinDist矩阵中
//若有新的minDist更新，重新计算meanMinDist
if (runs < N && runs > 2)
{
*meanMinDist_Pt = ((((float)(runs - 1)) * (*meanMinDist_Pt)) + (float)minDist) / ((float)runs);
}
else if (runs < N && runs == 2)
{
*meanMinDist_Pt = (float)minDist;
}
//1. update model
if (runs == N)
{
//Update current pixel
//check if random numer is smaller than ratio
//更新当前背景模型
if (randomT.at(entry) < ratio)
{
// replace randomly chosen sample
int rand = randomN.at(entry + 1); //randomGenerator.uniform((int)0,(int)N-1);
//替换梯度，灰度值
for (int z = 0; z < chans; ++z)
{
B_Mag_Pts.at(z).at(rand)[i] = (float)*currentFeaturesM_Pt.at(z);
B_Col_Pts.at(z).at(rand)[i] = (uchar)*currentFeaturesC_Pt.at(z);
}
//更新meanMinDist
*meanMinDist_Pt = ((((float)(N - 1)) * (*meanMinDist_Pt)) + (float)minDist) / ((float)N);
}
//邻域更新
//Update neighboring pixel model
if (randomTN.at(entry) < ratio)
{
//choose neighboring pixel randomly
int xNeigh = randomX.at(entry) + i;
int yNeigh = randomY.at(entry) + j;
//坐标测试
checkValid(&xNeigh, &yNeigh);
// replace randomly chosen sample
int rand = randomN.at(entry - 1);
//邻域更新
for (int z = 0; z < chans; ++z)
{
(backgroundModel.at(rand)).at(z).at<float>(yNeigh, xNeigh) = currentFeatures.at(z).at<float>(yNeigh, xNeigh);
(backgroundModel.at(rand)).at(z + chans).at<uchar>(yNeigh, xNeigh) = currentFeatures.at(z + chans).at<uchar>(yNeigh, xNeigh);
}
}
}
}
else //前景
{
// store pixel as foreground
*resultMap_Pt = 255;
//there is some movement
isMove = true;
}
//#######################//#######################//#######################//#######################
//control loops
//#######################//#######################//#######################//#######################
//update R
//更新R， 另一个版本的是在背景中更新R....【未解决】
decisionThresholdRegulator(actualR_Pt, meanMinDist_Pt);
//更新T
//update T
learningRateRegulator(actualT_Pt, meanMinDist_Pt, resultMap_Pt);
//#######################//#######################//#######################//#######################
//#######################//#######################//#######################//#######################
//jump to next pixel
++resultMap_Pt;
for (int z = 0; z < chans; ++z)
{
++currentFeaturesM_Pt.at(z);
++currentFeaturesC_Pt.at(z);
}
++meanMinDist_Pt;
++actualR_Pt;
++actualT_Pt;
}
}
resultMap->copyTo(*output);
//if there is no foreground -> no magnitudes fount
//-> initiate some low value to prevent diving through zero
//计算meanMag既公式5中的Im
double meanMag = sumMagnitude / (double)(glCounterFore + 1); //height*width);
if (meanMag > 20)
formerMeanMag = meanMag;
else
formerMeanMag = 20;
delete resultMap;
for (int z = 0; z < chans; ++z)
{
currentFeatures.at(z + chans).release();
currentFeatures.at(z).release();
}
return true;
}
//公式（3），更新Decision Threshold图(Mat)
//Pt指向当前R(xi), meanDist既Dmin均值
void PBAS::decisionThresholdRegulator(float* pt, float* meanDist)
{
//update R
double tempR = *pt;
double newThresh = (*meanDist)*R_scale;
//公式（3）
if (tempR < newThresh)
{
tempR += tempR * R_incdec;
}
else
{
tempR -= tempR * R_incdec;
}
//边界保护所有灰度阈值必须大于R_lower
if (tempR >= R_lower)
*pt = (float)tempR;
else
*pt = (float)R_lower;
}
//公式4，pt存储是当前learning rate 图(Mat), meanDist既Dmin均值，
void PBAS::learningRateRegulator(float* pt, float* meanDist, uchar* isFore)
{
//time update
double tempT = *pt;
//根据公式（4）更新
//若是背景
if ((int)*isFore < 128)
{
tempT -= T_inc / (*meanDist + 1.0);
}
else //若是前景
{
tempT += T_dec / (*meanDist + 1.0);
}
//最后更新的T(xi)必须在T_lower 与 T_upper之间
if (tempT > T_lower && tempT < T_upper)
*pt = (float)tempT;
}
//边界保护。
//因为模型Update的时候有一定概率更新邻域，而邻域位置是随机的，所以必须保证邻域位置有效
//既(x,y)坐标必须是有效数值
void PBAS::checkValid(int *x, int *y)
{
if (*x < 0)
{
*x = 0;
}
else if (*x >= width)
{
*x = width - 1;
}
if (*y < 0)
{
*y = 0;
}
else if (*y >= height)
{
*y = height - 1;
}
}
//初始化建立背景模型！！！！！！！！！！！！！
void PBAS::init(cv::Mat* input)
{
//对前N帧图像，依次计算每帧图像的Features，存入backgroundModel中
//最后runs = N，不再变化
if (runs < N)
{
std::vector<cv::Mat> init;
calculateFeatures(&init, input);
backgroundModel.push_back(init);
//以下可删
if (chans == 1)
{
init.at(0).release();
init.at(1).release();
}
else
{
init.at(0).release();
init.at(1).release();
init.at(2).release();
init.at(3).release();
init.at(4).release();
init.at(5).release();
}
init.clear();
//若是第一次，需要分配内存
if (runs == 0)
{
//存储公式3用到的Dmin均值，开始填充0
meanMinDist.create(input->size(), CV_32FC1);
meanMinDist.zeros(input->rows, input->cols, CV_32FC1);
//当前Decision Threshold矩阵，每个“像素”点赋值为R_lower,一般为18
actualR.create(input->rows, input->cols, CV_32FC1);
//当前Update Learning Rate矩阵，每个“像素”点赋值为T_init,一般为18
actualT.create(input->rows, input->cols, CV_32FC1);
float* ptRs, *ptTs; //, *ptM;
for (int rows = 0; rows < actualR.rows; ++rows)
{
ptRs = actualR.ptr<float>(rows);
ptTs = actualT.ptr<float>(rows);
for (int cols = 0; cols < actualR.cols; ++cols)
{
//依次赋值
ptRs[cols] = (float)R_lower;
ptTs[cols] = (float)T_init;
}
}
}
++runs;
}
}
//计算当前帧的Feature图
void PBAS::calculateFeatures(std::vector<cv::Mat>* feature, cv::Mat* inputImage)
{
//清空feature内容，因为建模完成后，每次计算的feature都存在currentFeature中，所以需要清空
if (!feature->empty())
feature->clear();
cv::Mat mag[3], dir;
//如果是彩色图像
if (inputImage->channels() == 3)
{
//拆分为R G B通道
std::vector<cv::Mat> rgbChannels(3);
cv::split(*inputImage, rgbChannels);
//依次对每个通道计算梯度，然后push到Feature中，此时feature存入了3个Mat，分别是
//R G B通道的梯度图像
for (int l = 0; l < 3; ++l)
{
cv::Sobel(rgbChannels.at(l), sobelX, CV_32F, 1, 0, 3, 1, 0.0);
cv::Sobel(rgbChannels.at(l), sobelY, CV_32F, 0, 1, 3, 1, 0.0);
// Compute the L2 norm and direction of the gradient
cv::cartToPolar(sobelX, sobelY, mag[l], dir, true);
feature->push_back(mag[l]);
sobelX.release();
sobelY.release();
}
//然后再把R G B通道的灰度值push到feature 中...
feature->push_back(rgbChannels.at(0));
feature->push_back(rgbChannels.at(1));
feature->push_back(rgbChannels.at(2));
rgbChannels.at(0).release();
rgbChannels.at(1).release();
rgbChannels.at(2).release();
//总结就是：
//feature这个vector<Mat> 最后存了6个Mat，分别是梯度，梯度，梯度，灰度，灰度，灰度
}
else
{
//灰度图像同理，
//只不过最后feature只存入2个Mat， 分别对应梯度图，灰度值
cv::Sobel(*inputImage, sobelX, CV_32F, 1, 0, 3, 1, 0.0);
cv::Sobel(*inputImage, sobelY, CV_32F, 0, 1, 3, 1, 0.0);
// Compute the L2 norm and direction of the gradient
cv::cartToPolar(sobelX, sobelY, mag[0], dir, true);
feature->push_back(mag[0]);
cv::Mat temp;
inputImage->copyTo(temp);
feature->push_back(temp);
temp.release();
}
mag[0].release();
mag[1].release();
mag[2].release();
dir.release();
}
//以下都是设置参数函数，不说了................
void PBAS::setN(int temp)
{
N = temp;
newInitialization(); //因为N关系到随机数产生器
}
void PBAS::setRaute_min(int temp)
{
Raute_min = temp;
}
void PBAS::setR_lower(double temp)
{
R_lower = temp;
}
void PBAS::setR_incdec(double temp)
{
R_incdec = temp;
}
void PBAS::setR_scale(double temp)
{
R_scale = temp;
}
void PBAS::setT_init(double temp)
{
T_init = temp;
}
void PBAS::setT_lower(double temp)
{
T_lower = temp;
}
void PBAS::setT_upper(double temp)
{
T_upper = temp;
newInitialization(); //因为 T_upper关系到随机数产生器
}
void PBAS::setT_dec(double temp)
{
T_dec = temp;
}
void PBAS::setT_inc(double temp)
{
T_inc = temp;
}
void PBAS::setAlpha(double temp)
{
alpha = temp;
}
void PBAS::setBeta(double temp)
{
beta = temp;
}
bool PBAS::isMovement()
{
return isMove;
}
//cv::Mat* PBAS::getR1_xi()
//{
// return &actualR;
//}
//
//cv::Mat* PBAS::getT_xi()
//{
// return &actualT;
//}
main.cpp
#include <iostream>
#include "pbas.h"
using namespace cv;
using namespace std;
int main(int argc, char **argv)
{
std::cout << "Using OpenCV " << CV_MAJOR_VERSION << "." << CV_MINOR_VERSION << "." << CV_SUBMINOR_VERSION << std::endl;
/* Open video file */
VideoCapture capture(0);
if (!capture.isOpened())
{
std::cerr << "Cannot open video!" << std::endl;
return 1;
}
PBAS pbas;
std::cout << "Press 'q' to quit..." << std::endl;
int key = 0;
Mat img_input;
while (key != 'q')
{
capture >> img_input;
cv::imshow("Input", img_input);
cv::GaussianBlur(img_input, img_input, cv::Size(5, 5), 1.5);
cv::Mat img_mask;
double t = (double)cv::getTickCount();
pbas.process(&img_input, &img_mask);
t = (double)cv::getTickCount() - t;
int fps = (int)(cv::getTickFrequency() / t);
std::cout << fps << std::endl;
cv::medianBlur(img_mask, img_mask, 5);
imshow("img_mask", img_mask);
key = cvWaitKey(1);
}
return 0;
}
6730
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
