提交 69779309 编写于 作者: V Vladislav Vinogradov

refactored GMG algorithm

上级 b8f0d1a0
......@@ -91,6 +91,20 @@ CV_EXPORTS Ptr<gpu::BackgroundSubtractorMOG2>
createBackgroundSubtractorMOG2(int history = 500, double varThreshold = 16,
bool detectShadows = true);
////////////////////////////////////////////////////
// GMG
class CV_EXPORTS BackgroundSubtractorGMG : public cv::BackgroundSubtractorGMG
{
public:
using cv::BackgroundSubtractorGMG::apply;
virtual void apply(InputArray image, OutputArray fgmask, double learningRate, Stream& stream) = 0;
};
CV_EXPORTS Ptr<gpu::BackgroundSubtractorGMG>
createBackgroundSubtractorGMG(int initializationFrames = 120, double decisionThreshold = 0.8);
......@@ -161,77 +175,6 @@ private:
std::auto_ptr<Impl> impl_;
};
/**
* Background Subtractor module. Takes a series of images and returns a sequence of mask (8UC1)
* images of the same size, where 255 indicates Foreground and 0 represents Background.
* This class implements an algorithm described in "Visual Tracking of Human Visitors under
* Variable-Lighting Conditions for a Responsive Audio Art Installation," A. Godbehere,
* A. Matsukawa, K. Goldberg, American Control Conference, Montreal, June 2012.
*/
class CV_EXPORTS GMG_GPU
{
public:
GMG_GPU();
/**
* Validate parameters and set up data structures for appropriate frame size.
* @param frameSize Input frame size
* @param min Minimum value taken on by pixels in image sequence. Usually 0
* @param max Maximum value taken on by pixels in image sequence. e.g. 1.0 or 255
*/
void initialize(Size frameSize, float min = 0.0f, float max = 255.0f);
/**
* Performs single-frame background subtraction and builds up a statistical background image
* model.
* @param frame Input frame
* @param fgmask Output mask image representing foreground and background pixels
* @param stream Stream for the asynchronous version
*/
void operator ()(const GpuMat& frame, GpuMat& fgmask, float learningRate = -1.0f, Stream& stream = Stream::Null());
//! Releases all inner buffers
void release();
//! Total number of distinct colors to maintain in histogram.
int maxFeatures;
//! Set between 0.0 and 1.0, determines how quickly features are "forgotten" from histograms.
float learningRate;
//! Number of frames of video to use to initialize histograms.
int numInitializationFrames;
//! Number of discrete levels in each channel to be used in histograms.
int quantizationLevels;
//! Prior probability that any given pixel is a background pixel. A sensitivity parameter.
float backgroundPrior;
//! Value above which pixel is determined to be FG.
float decisionThreshold;
//! Smoothing radius, in pixels, for cleaning up FG image.
int smoothingRadius;
//! Perform background model update.
bool updateBackgroundModel;
private:
float maxVal_, minVal_;
Size frameSize_;
int frameNum_;
GpuMat nfeatures_;
GpuMat colors_;
GpuMat weights_;
Ptr<gpu::Filter> boxFilter_;
GpuMat buf_;
};
}} // namespace cv { namespace gpu {
#endif /* __OPENCV_GPUBGSEGM_HPP__ */
......@@ -462,10 +462,10 @@ PERF_TEST_P(Video_Cn_MaxFeatures, GMG,
cv::gpu::GpuMat d_frame(frame);
cv::gpu::GpuMat foreground;
cv::gpu::GMG_GPU d_gmg;
d_gmg.maxFeatures = maxFeatures;
cv::Ptr<cv::BackgroundSubtractorGMG> d_gmg = cv::gpu::createBackgroundSubtractorGMG();
d_gmg->setMaxFeatures(maxFeatures);
d_gmg(d_frame, foreground);
d_gmg->apply(d_frame, foreground);
for (int i = 0; i < 150; ++i)
{
......@@ -490,7 +490,7 @@ PERF_TEST_P(Video_Cn_MaxFeatures, GMG,
d_frame.upload(frame);
startTimer(); next();
d_gmg(d_frame, foreground);
d_gmg->apply(d_frame, foreground);
stopTimer();
}
......@@ -501,9 +501,8 @@ PERF_TEST_P(Video_Cn_MaxFeatures, GMG,
cv::Mat foreground;
cv::Mat zeros(frame.size(), CV_8UC1, cv::Scalar::all(0));
cv::Ptr<cv::BackgroundSubtractor> gmg = cv::createBackgroundSubtractorGMG();
gmg->set("maxFeatures", maxFeatures);
//gmg.initialize(frame.size(), 0.0, 255.0);
cv::Ptr<cv::BackgroundSubtractorGMG> gmg = cv::createBackgroundSubtractorGMG();
gmg->setMaxFeatures(maxFeatures);
gmg->apply(frame, foreground);
......
......@@ -47,7 +47,7 @@
#include "opencv2/core/cuda/limits.hpp"
namespace cv { namespace gpu { namespace cudev {
namespace bgfg_gmg
namespace gmg
{
__constant__ int c_width;
__constant__ int c_height;
......
......@@ -42,17 +42,17 @@
#include "precomp.hpp"
#if !defined HAVE_CUDA || defined(CUDA_DISABLER)
using namespace cv;
using namespace cv::gpu;
cv::gpu::GMG_GPU::GMG_GPU() { throw_no_cuda(); }
void cv::gpu::GMG_GPU::initialize(cv::Size, float, float) { throw_no_cuda(); }
void cv::gpu::GMG_GPU::operator ()(const cv::gpu::GpuMat&, cv::gpu::GpuMat&, float, cv::gpu::Stream&) { throw_no_cuda(); }
void cv::gpu::GMG_GPU::release() {}
#if !defined HAVE_CUDA || defined(CUDA_DISABLER) || !defined(HAVE_OPENCV_GPUFILTERS)
Ptr<gpu::BackgroundSubtractorGMG> cv::gpu::createBackgroundSubtractorGMG(int, double) { throw_no_cuda(); return Ptr<gpu::BackgroundSubtractorGMG>(); }
#else
namespace cv { namespace gpu { namespace cudev {
namespace bgfg_gmg
namespace gmg
{
void loadConstants(int width, int height, float minVal, float maxVal, int quantizationLevels, float backgroundPrior,
float decisionThreshold, int maxFeatures, int numInitializationFrames);
......@@ -63,103 +63,209 @@ namespace cv { namespace gpu { namespace cudev {
}
}}}
cv::gpu::GMG_GPU::GMG_GPU()
namespace
{
maxFeatures = 64;
learningRate = 0.025f;
numInitializationFrames = 120;
quantizationLevels = 16;
backgroundPrior = 0.8f;
decisionThreshold = 0.8f;
smoothingRadius = 7;
updateBackgroundModel = true;
}
class GMGImpl : public gpu::BackgroundSubtractorGMG
{
public:
GMGImpl(int initializationFrames, double decisionThreshold);
void cv::gpu::GMG_GPU::initialize(cv::Size frameSize, float min, float max)
{
using namespace cv::gpu::cudev::bgfg_gmg;
void apply(InputArray image, OutputArray fgmask, double learningRate=-1);
void apply(InputArray image, OutputArray fgmask, double learningRate, Stream& stream);
CV_Assert(min < max);
CV_Assert(maxFeatures > 0);
CV_Assert(learningRate >= 0.0f && learningRate <= 1.0f);
CV_Assert(numInitializationFrames >= 1);
CV_Assert(quantizationLevels >= 1 && quantizationLevels <= 255);
CV_Assert(backgroundPrior >= 0.0f && backgroundPrior <= 1.0f);
void getBackgroundImage(OutputArray backgroundImage) const;
minVal_ = min;
maxVal_ = max;
int getMaxFeatures() const { return maxFeatures_; }
void setMaxFeatures(int maxFeatures) { maxFeatures_ = maxFeatures; }
frameSize_ = frameSize;
double getDefaultLearningRate() const { return learningRate_; }
void setDefaultLearningRate(double lr) { learningRate_ = (float) lr; }
frameNum_ = 0;
int getNumFrames() const { return numInitializationFrames_; }
void setNumFrames(int nframes) { numInitializationFrames_ = nframes; }
nfeatures_.create(frameSize_, CV_32SC1);
colors_.create(maxFeatures * frameSize_.height, frameSize_.width, CV_32SC1);
weights_.create(maxFeatures * frameSize_.height, frameSize_.width, CV_32FC1);
int getQuantizationLevels() const { return quantizationLevels_; }
void setQuantizationLevels(int nlevels) { quantizationLevels_ = nlevels; }
nfeatures_.setTo(cv::Scalar::all(0));
double getBackgroundPrior() const { return backgroundPrior_; }
void setBackgroundPrior(double bgprior) { backgroundPrior_ = (float) bgprior; }
if (smoothingRadius > 0)
boxFilter_ = cv::gpu::createBoxFilter(CV_8UC1, -1, cv::Size(smoothingRadius, smoothingRadius));
int getSmoothingRadius() const { return smoothingRadius_; }
void setSmoothingRadius(int radius) { smoothingRadius_ = radius; }
loadConstants(frameSize_.width, frameSize_.height, minVal_, maxVal_, quantizationLevels, backgroundPrior, decisionThreshold, maxFeatures, numInitializationFrames);
}
double getDecisionThreshold() const { return decisionThreshold_; }
void setDecisionThreshold(double thresh) { decisionThreshold_ = (float) thresh; }
void cv::gpu::GMG_GPU::operator ()(const cv::gpu::GpuMat& frame, cv::gpu::GpuMat& fgmask, float newLearningRate, cv::gpu::Stream& stream)
{
using namespace cv::gpu::cudev::bgfg_gmg;
bool getUpdateBackgroundModel() const { return updateBackgroundModel_; }
void setUpdateBackgroundModel(bool update) { updateBackgroundModel_ = update; }
typedef void (*func_t)(PtrStepSzb frame, PtrStepb fgmask, PtrStepSzi colors, PtrStepf weights, PtrStepi nfeatures,
int frameNum, float learningRate, bool updateBackgroundModel, cudaStream_t stream);
static const func_t funcs[6][4] =
{
{update_gpu<uchar>, 0, update_gpu<uchar3>, update_gpu<uchar4>},
{0,0,0,0},
{update_gpu<ushort>, 0, update_gpu<ushort3>, update_gpu<ushort4>},
{0,0,0,0},
{0,0,0,0},
{update_gpu<float>, 0, update_gpu<float3>, update_gpu<float4>}
double getMinVal() const { return minVal_; }
void setMinVal(double val) { minVal_ = (float) val; }
double getMaxVal() const { return maxVal_; }
void setMaxVal(double val) { maxVal_ = (float) val; }
private:
void initialize(Size frameSize, float min, float max);
//! Total number of distinct colors to maintain in histogram.
int maxFeatures_;
//! Set between 0.0 and 1.0, determines how quickly features are "forgotten" from histograms.
float learningRate_;
//! Number of frames of video to use to initialize histograms.
int numInitializationFrames_;
//! Number of discrete levels in each channel to be used in histograms.
int quantizationLevels_;
//! Prior probability that any given pixel is a background pixel. A sensitivity parameter.
float backgroundPrior_;
//! Smoothing radius, in pixels, for cleaning up FG image.
int smoothingRadius_;
//! Value above which pixel is determined to be FG.
float decisionThreshold_;
//! Perform background model update.
bool updateBackgroundModel_;
float minVal_, maxVal_;
Size frameSize_;
int frameNum_;
GpuMat nfeatures_;
GpuMat colors_;
GpuMat weights_;
Ptr<gpu::Filter> boxFilter_;
GpuMat buf_;
};
CV_Assert(frame.depth() == CV_8U || frame.depth() == CV_16U || frame.depth() == CV_32F);
CV_Assert(frame.channels() == 1 || frame.channels() == 3 || frame.channels() == 4);
GMGImpl::GMGImpl(int initializationFrames, double decisionThreshold)
{
maxFeatures_ = 64;
learningRate_ = 0.025f;
numInitializationFrames_ = initializationFrames;
quantizationLevels_ = 16;
backgroundPrior_ = 0.8f;
decisionThreshold_ = (float) decisionThreshold;
smoothingRadius_ = 7;
updateBackgroundModel_ = true;
minVal_ = maxVal_ = 0;
}
if (newLearningRate != -1.0f)
void GMGImpl::apply(InputArray image, OutputArray fgmask, double learningRate)
{
CV_Assert(newLearningRate >= 0.0f && newLearningRate <= 1.0f);
learningRate = newLearningRate;
apply(image, fgmask, learningRate, Stream::Null());
}
if (frame.size() != frameSize_)
initialize(frame.size(), 0.0f, frame.depth() == CV_8U ? 255.0f : frame.depth() == CV_16U ? std::numeric_limits<ushort>::max() : 1.0f);
void GMGImpl::apply(InputArray _frame, OutputArray _fgmask, double newLearningRate, Stream& stream)
{
using namespace cv::gpu::cudev::gmg;
typedef void (*func_t)(PtrStepSzb frame, PtrStepb fgmask, PtrStepSzi colors, PtrStepf weights, PtrStepi nfeatures,
int frameNum, float learningRate, bool updateBackgroundModel, cudaStream_t stream);
static const func_t funcs[6][4] =
{
{update_gpu<uchar>, 0, update_gpu<uchar3>, update_gpu<uchar4>},
{0,0,0,0},
{update_gpu<ushort>, 0, update_gpu<ushort3>, update_gpu<ushort4>},
{0,0,0,0},
{0,0,0,0},
{update_gpu<float>, 0, update_gpu<float3>, update_gpu<float4>}
};
GpuMat frame = _frame.getGpuMat();
CV_Assert( frame.depth() == CV_8U || frame.depth() == CV_16U || frame.depth() == CV_32F );
CV_Assert( frame.channels() == 1 || frame.channels() == 3 || frame.channels() == 4 );
if (newLearningRate != -1.0)
{
CV_Assert( newLearningRate >= 0.0 && newLearningRate <= 1.0 );
learningRate_ = (float) newLearningRate;
}
if (frame.size() != frameSize_)
{
double minVal = minVal_;
double maxVal = maxVal_;
fgmask.create(frameSize_, CV_8UC1);
fgmask.setTo(cv::Scalar::all(0), stream);
if (minVal_ == 0 && maxVal_ == 0)
{
minVal = 0;
maxVal = frame.depth() == CV_8U ? 255.0 : frame.depth() == CV_16U ? std::numeric_limits<ushort>::max() : 1.0;
}
funcs[frame.depth()][frame.channels() - 1](frame, fgmask, colors_, weights_, nfeatures_, frameNum_, learningRate, updateBackgroundModel, cv::gpu::StreamAccessor::getStream(stream));
initialize(frame.size(), (float) minVal, (float) maxVal);
}
_fgmask.create(frameSize_, CV_8UC1);
GpuMat fgmask = _fgmask.getGpuMat();
fgmask.setTo(Scalar::all(0), stream);
funcs[frame.depth()][frame.channels() - 1](frame, fgmask, colors_, weights_, nfeatures_, frameNum_,
learningRate_, updateBackgroundModel_, StreamAccessor::getStream(stream));
// medianBlur
if (smoothingRadius_ > 0)
{
boxFilter_->apply(fgmask, buf_, stream);
const int minCount = (smoothingRadius_ * smoothingRadius_ + 1) / 2;
const double thresh = 255.0 * minCount / (smoothingRadius_ * smoothingRadius_);
gpu::threshold(buf_, fgmask, thresh, 255.0, THRESH_BINARY, stream);
}
// keep track of how many frames we have processed
++frameNum_;
}
// medianBlur
if (smoothingRadius > 0)
void GMGImpl::getBackgroundImage(OutputArray backgroundImage) const
{
boxFilter_->apply(fgmask, buf_, stream);
int minCount = (smoothingRadius * smoothingRadius + 1) / 2;
double thresh = 255.0 * minCount / (smoothingRadius * smoothingRadius);
cv::gpu::threshold(buf_, fgmask, thresh, 255.0, cv::THRESH_BINARY, stream);
(void) backgroundImage;
CV_Error(Error::StsNotImplemented, "Not implemented");
}
// keep track of how many frames we have processed
++frameNum_;
void GMGImpl::initialize(Size frameSize, float min, float max)
{
using namespace cv::gpu::cudev::gmg;
CV_Assert( maxFeatures_ > 0 );
CV_Assert( learningRate_ >= 0.0f && learningRate_ <= 1.0f);
CV_Assert( numInitializationFrames_ >= 1);
CV_Assert( quantizationLevels_ >= 1 && quantizationLevels_ <= 255);
CV_Assert( backgroundPrior_ >= 0.0f && backgroundPrior_ <= 1.0f);
minVal_ = min;
maxVal_ = max;
CV_Assert( minVal_ < maxVal_ );
frameSize_ = frameSize;
frameNum_ = 0;
nfeatures_.create(frameSize_, CV_32SC1);
colors_.create(maxFeatures_ * frameSize_.height, frameSize_.width, CV_32SC1);
weights_.create(maxFeatures_ * frameSize_.height, frameSize_.width, CV_32FC1);
nfeatures_.setTo(Scalar::all(0));
if (smoothingRadius_ > 0)
boxFilter_ = gpu::createBoxFilter(CV_8UC1, -1, Size(smoothingRadius_, smoothingRadius_));
loadConstants(frameSize_.width, frameSize_.height, minVal_, maxVal_,
quantizationLevels_, backgroundPrior_, decisionThreshold_, maxFeatures_, numInitializationFrames_);
}
}
void cv::gpu::GMG_GPU::release()
Ptr<gpu::BackgroundSubtractorGMG> cv::gpu::createBackgroundSubtractorGMG(int initializationFrames, double decisionThreshold)
{
frameSize_ = Size();
nfeatures_.release();
colors_.release();
weights_.release();
boxFilter_.release();
buf_.release();
return new GMGImpl(initializationFrames, decisionThreshold);
}
#endif
......@@ -52,4 +52,6 @@
#include "opencv2/core/private.gpu.hpp"
#include "opencv2/opencv_modules.hpp"
#endif /* __OPENCV_PRECOMP_H__ */
......@@ -372,16 +372,15 @@ GPU_TEST_P(GMG, Accuracy)
cv::Mat frame = randomMat(size, type, 0, 100);
cv::gpu::GpuMat d_frame = loadMat(frame, useRoi);
cv::gpu::GMG_GPU gmg;
gmg.numInitializationFrames = 5;
gmg.smoothingRadius = 0;
gmg.initialize(d_frame.size(), 0, 255);
cv::Ptr<cv::BackgroundSubtractorGMG> gmg = cv::gpu::createBackgroundSubtractorGMG();
gmg->setNumFrames(5);
gmg->setSmoothingRadius(0);
cv::gpu::GpuMat d_fgmask = createMat(size, CV_8UC1, useRoi);
for (int i = 0; i < gmg.numInitializationFrames; ++i)
for (int i = 0; i < gmg->getNumFrames(); ++i)
{
gmg(d_frame, d_fgmask);
gmg->apply(d_frame, d_fgmask);
// fgmask should be entirely background during training
ASSERT_MAT_NEAR(zeros, d_fgmask, 0);
......@@ -389,7 +388,7 @@ GPU_TEST_P(GMG, Accuracy)
frame = randomMat(size, type, 160, 255);
d_frame = loadMat(frame, useRoi);
gmg(d_frame, d_fgmask);
gmg->apply(d_frame, d_fgmask);
// now fgmask should be entirely foreground
ASSERT_MAT_NEAR(fullfg, d_fgmask, 0);
......
......@@ -18,10 +18,10 @@ using namespace cv::gpu;
enum Method
{
FGD_STAT,
MOG,
MOG2,
GMG
GMG,
FGD_STAT
};
int main(int argc, const char** argv)
......@@ -29,7 +29,7 @@ int main(int argc, const char** argv)
cv::CommandLineParser cmd(argc, argv,
"{ c camera | | use camera }"
"{ f file | 768x576.avi | input video file }"
"{ m method | mog | method (fgd, mog, mog2, gmg) }"
"{ m method | mog | method (mog, mog2, gmg, fgd) }"
"{ h help | | print help message }");
if (cmd.has("help") || !cmd.check())
......@@ -43,18 +43,18 @@ int main(int argc, const char** argv)
string file = cmd.get<string>("file");
string method = cmd.get<string>("method");
if (method != "fgd"
&& method != "mog"
if (method != "mog"
&& method != "mog2"
&& method != "gmg")
&& method != "gmg"
&& method != "fgd")
{
cerr << "Incorrect method" << endl;
return -1;
}
Method m = method == "fgd" ? FGD_STAT :
method == "mog" ? MOG :
Method m = method == "mog" ? MOG :
method == "mog2" ? MOG2 :
method == "fgd" ? FGD_STAT :
GMG;
VideoCapture cap;
......@@ -75,11 +75,10 @@ int main(int argc, const char** argv)
GpuMat d_frame(frame);
Ptr<BackgroundSubtractor> mog = gpu::createBackgroundSubtractorMOG();
Ptr<BackgroundSubtractor> mog2 = gpu::createBackgroundSubtractorMOG2();
Ptr<BackgroundSubtractor> gmg = gpu::createBackgroundSubtractorGMG(40);
FGDStatModel fgd_stat;
cv::Ptr<cv::BackgroundSubtractor> mog = cv::gpu::createBackgroundSubtractorMOG();
cv::Ptr<cv::BackgroundSubtractor> mog2 = cv::gpu::createBackgroundSubtractorMOG2();
GMG_GPU gmg;
gmg.numInitializationFrames = 40;
GpuMat d_fgmask;
GpuMat d_fgimg;
......@@ -91,10 +90,6 @@ int main(int argc, const char** argv)
switch (m)
{
case FGD_STAT:
fgd_stat.create(d_frame);
break;
case MOG:
mog->apply(d_frame, d_fgmask, 0.01);
break;
......@@ -104,7 +99,11 @@ int main(int argc, const char** argv)
break;
case GMG:
gmg.initialize(d_frame.size());
gmg->apply(d_frame, d_fgmask);
break;
case FGD_STAT:
fgd_stat.create(d_frame);
break;
}
......@@ -128,12 +127,6 @@ int main(int argc, const char** argv)
//update the model
switch (m)
{
case FGD_STAT:
fgd_stat.update(d_frame);
d_fgmask = fgd_stat.foreground;
d_bgimg = fgd_stat.background;
break;
case MOG:
mog->apply(d_frame, d_fgmask, 0.01);
mog->getBackgroundImage(d_bgimg);
......@@ -145,7 +138,13 @@ int main(int argc, const char** argv)
break;
case GMG:
gmg(d_frame, d_fgmask);
gmg->apply(d_frame, d_fgmask);
break;
case FGD_STAT:
fgd_stat.update(d_frame);
d_fgmask = fgd_stat.foreground;
d_bgimg = fgd_stat.background;
break;
}
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册