void update_cur_bbox_vec(std::vector<bbox_t> _cur_bbox_vec)
{
cur_bbox_vec = _cur_bbox_vec;
good_bbox_vec_flags = std::vector<bool>(cur_bbox_vec.size(), true);
cv::Mat prev_pts, cur_pts_flow_cpu;
for (auto &i : cur_bbox_vec) {
float x_center = (i.x + i.w / 2.0F);
float y_center = (i.y + i.h / 2.0F);
prev_pts.push_back(cv::Point2f(x_center, y_center));
}
if (prev_pts.rows == 0)
prev_pts_flow_cpu = cv::Mat();
else
cv::transpose(prev_pts, prev_pts_flow_cpu);
if (prev_pts_flow_gpu.cols < prev_pts_flow_cpu.cols) {
prev_pts_flow_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), prev_pts_flow_cpu.type());
cur_pts_flow_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), prev_pts_flow_cpu.type());
status_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), CV_8UC1);
err_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), CV_32FC1);
}
prev_pts_flow_gpu.upload(cv::Mat(prev_pts_flow_cpu), stream);
}
void CGLUtil::gpuMapRGBResources(const cv::cuda::GpuMat& cvgmRGBs_){
int nPyrLevel_ = getLevel( cvgmRGBs_.cols );
// map OpenGL buffer object for writing from CUDA
void *pDev;
cudaGraphicsMapResources(1, &_apResourceRGBVBO[nPyrLevel_], 0);
size_t nSize;
cudaGraphicsResourceGetMappedPointer((void **)&pDev, &nSize, _apResourceRGBVBO[nPyrLevel_] );
cv::cuda::GpuMat cvgmRGBs(cvgmRGBs_.size(),CV_8UC3,pDev);
cvgmRGBs_.copyTo(cvgmRGBs);
cudaGraphicsUnmapResources(1, &_apResourceRGBVBO[nPyrLevel_], 0);
// render from the vbo
glBindBuffer(GL_ARRAY_BUFFER, _auRGBVBO[nPyrLevel_]);
glColorPointer(3, GL_UNSIGNED_BYTE, 0, 0);
glEnableClientState(GL_COLOR_ARRAY);//you cant move glEnableClientState infront of cuda GraphicsMapResources, otherwise, you will have weird problem
//glDrawArrays(GL_POINTS, 0, btl::kinect::__aKinectWxH[usPyrLevel_] );
//glDisableClientState(GL_COLOR_ARRAY);
//glBindBuffer( GL_ARRAY_BUFFER, 0 );
}
void CGLUtil::gpuMapNlResources(const cv::cuda::GpuMat& cvgmNls_){
int nPyrLevel_ = getLevel( cvgmNls_.cols );
// map OpenGL buffer object for writing from CUDA
void *pDev;
cudaGraphicsMapResources(1, &_apResourceNlVBO[nPyrLevel_], 0);
size_t nSize;
cudaGraphicsResourceGetMappedPointer((void **)&pDev, &nSize, _apResourceNlVBO[nPyrLevel_] );
cv::cuda::GpuMat cvgmNls(cvgmNls_.size(),CV_32FC3,pDev);
cvgmNls_.copyTo(cvgmNls);
cudaGraphicsUnmapResources(1, &_apResourceNlVBO[nPyrLevel_], 0);
// render from the vbo
glBindBuffer(GL_ARRAY_BUFFER, _auNlVBO[nPyrLevel_]);
glNormalPointer(GL_FLOAT, 12, 0); //12 is the stride = the number of bytes occupied by each normal
glEnableClientState(GL_NORMAL_ARRAY);//you cant move glEnableClientState infront of cuda GraphicsMapResources, otherwise, you will have weird problem
//glColor3f(1.0, 0.0, 0.0);
//glDrawArrays(GL_POINTS, 0, btl::kinect::__aKinectWxH[usPyrLevel_] );
//glDisableClientState(GL_NORMAL_ARRAY);
//glBindBuffer( GL_ARRAY_BUFFER, 0 );
}
void CGLUtil::gpuMapPtResources(const cv::cuda::GpuMat& cvgmPts_){
int nPyrLevel_ = getLevel( cvgmPts_.cols );
// map OpenGL buffer object for writing from CUDA
void *pDev;
cudaGraphicsMapResources(1, &_apResourcePtVBO[nPyrLevel_], 0);
size_t nSize;
cudaGraphicsResourceGetMappedPointer((void **)&pDev, &nSize, _apResourcePtVBO[nPyrLevel_] );
cv::cuda::GpuMat cvgmPts( cvgmPts_.size(),CV_32FC3,pDev );
cvgmPts_.copyTo(cvgmPts); // the operation of the Buffer must be done before cudaGraphicsUnmapResources(), otherwise, the buffer will affects each other
cudaGraphicsUnmapResources(1, &_apResourcePtVBO[nPyrLevel_], 0);
// render from the vbo
glBindBuffer(GL_ARRAY_BUFFER, _auPtVBO[nPyrLevel_]);
glVertexPointer(3, GL_FLOAT, 0, 0);
glEnableClientState(GL_VERTEX_ARRAY);//you cant move glEnableClientState in front of cuda GraphicsMapResources, otherwise, you will have weird problem
//glColor3f(1.0, 0.0, 0.0);
//glDrawArrays(GL_POINTS, 0, btl::kinect::__aKinectWxH[usPyrLevel_] );
//glDisableClientState(GL_VERTEX_ARRAY);
//glBindBuffer( GL_ARRAY_BUFFER, 0 );
}
void enqueueSend(cv::cuda::GpuMat& m, cv::cuda::Stream& strm)
{
m.download(image_data_, strm);
strm.enqueueHostCallback(
[](int status, void *userData)
{
(void)status;
static_cast<GPUSender *>(userData)->send();
},
(void *)this);
}
void _SSD::Preprocess(const cv::cuda::GpuMat& img, std::vector<cv::cuda::GpuMat>* input_channels)
{
/* Convert the input image to the input image format of the network. */
cv::cuda::GpuMat sample;
if (img.channels() == 3 && num_channels_ == 1)
cv::cuda::cvtColor(img, sample, cv::COLOR_BGR2GRAY);
else if (img.channels() == 4 && num_channels_ == 1)
cv::cuda::cvtColor(img, sample, cv::COLOR_BGRA2GRAY);
else if (img.channels() == 4 && num_channels_ == 3)
cv::cuda::cvtColor(img, sample, cv::COLOR_BGRA2BGR);
else if (img.channels() == 1 && num_channels_ == 3)
cv::cuda::cvtColor(img, sample, cv::COLOR_GRAY2BGR);
else
sample = img;
cv::cuda::GpuMat sample_resized;
if (sample.size() != input_geometry_)
cv::cuda::resize(sample, sample_resized, input_geometry_);
else
sample_resized = sample;
cv::cuda::GpuMat sample_float;
if (num_channels_ == 3)
sample_resized.convertTo(sample_float, CV_32FC3);
else
sample_resized.convertTo(sample_float, CV_32FC1);
cv::cuda::GpuMat sample_normalized;
// cv::subtract(sample_float, mean_, sample_normalized);
sample_normalized = sample_float;
/* This operation will write the separate BGR planes directly to the
* input layer of the network because it is wrapped by the cv::Mat
* objects in input_channels. */
cv::cuda::split(sample_normalized, *input_channels);
// CHECK(reinterpret_cast<float*>(input_channels->at(0).data)
// == net_->input_blobs()[0]->cpu_data())
// << "Input channels are not wrapping the input layer of the network.";
}
void update_tracking_flow(cv::Mat src_mat, std::vector<bbox_t> _cur_bbox_vec)
{
int const old_gpu_id = cv::cuda::getDevice();
if (old_gpu_id != gpu_id)
cv::cuda::setDevice(gpu_id);
if (src_mat.channels() == 3) {
if (src_mat_gpu.cols == 0) {
src_mat_gpu = cv::cuda::GpuMat(src_mat.size(), src_mat.type());
src_grey_gpu = cv::cuda::GpuMat(src_mat.size(), CV_8UC1);
}
update_cur_bbox_vec(_cur_bbox_vec);
//src_grey_gpu.upload(src_mat, stream); // use BGR
src_mat_gpu.upload(src_mat, stream);
cv::cuda::cvtColor(src_mat_gpu, src_grey_gpu, CV_BGR2GRAY, 1, stream);
}
if (old_gpu_id != gpu_id)
cv::cuda::setDevice(old_gpu_id);
}
bool Optimizer::optimizeA(const cv::cuda::GpuMat _d,cv::cuda::GpuMat _a){
using namespace cv::cuda::device::dtam_optimizer;
localStream = cv::cuda::StreamAccessor::getStream(cvStream);
this->_a=_a;
Mat tmp(cv.rows,cv.cols,CV_32FC1);
bool doneOptimizing = theta <= thetaMin;
int layerStep = cv.rows * cv.cols;
float* d = (float*) _d.data;
float* a = (float*) _a.data;
loadConstants(cv.rows, cv.cols, cv.layers, layerStep, a, d, cv.data, (float*)cv.lo.data,
(float*)cv.hi.data, (float*)cv.loInd.data);
minimizeACaller ( cv.data, a, d, cv.layers, theta,lambda);
theta*=thetaStep;
if (doneOptimizing){
stableDepthReady=Ptr<char>((char*)(new cudaEvent_t));
cudaEventCreate((cudaEvent_t*)(char*)stableDepthReady,cudaEventBlockingSync);
_a.convertTo(stableDepth,CV_32FC1,cv.depthStep,cv.far,cvStream);
cudaEventRecord(*(cudaEvent_t*)(char*)stableDepthReady,localStream);
stableDepthEnqueued = 1;
}
return doneOptimizing;
}
GLuint CGLUtil::gpuMapRgb2PixelBufferObj(const cv::cuda::GpuMat& cvgmRGB_ ){
//http://rickarkin.blogspot.co.uk/2012/03/use-pbo-to-share-buffer-between-cuda.html
int nPyrLevel_ = getLevel( cvgmRGB_.cols );
GLuint uTexture;
// map OpenGL buffer object for writing from CUDA
if (cvgmRGB_.channels() == 3) {
uTexture = _auTexture[nPyrLevel_];
void *pDev;
cudaSafeCall( cudaGraphicsMapResources(1, &_apResourceRGBPxielBO[nPyrLevel_], 0));
size_t nSize;
cudaSafeCall( cudaGraphicsResourceGetMappedPointer((void **)&pDev, &nSize , _apResourceRGBPxielBO[nPyrLevel_]));
cv::cuda::GpuMat cvgmRGBA( cvgmRGB_.size(), CV_8UC3, pDev);
cvgmRGB_.copyTo(cvgmRGBA);
cudaSafeCall( cudaGraphicsUnmapResources(1, &_apResourceRGBPxielBO[nPyrLevel_], 0) );
//texture mapping
glBindTexture( GL_TEXTURE_2D, uTexture);
glBindBuffer ( GL_PIXEL_UNPACK_BUFFER_ARB, _auRGBPixelBO[nPyrLevel_]);
glTexImage2D( GL_TEXTURE_2D, 0, GL_RGB, cvgmRGB_.cols, cvgmRGB_.rows, 0, GL_RGB, GL_UNSIGNED_BYTE, NULL);
errorDetectorGL();
glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
glBindTexture(GL_TEXTURE_2D, 0);
}
else if (cvgmRGB_.channels()==1) {
uTexture = _auGrayTexture[nPyrLevel_];
void *pDev;
cudaSafeCall( cudaGraphicsMapResources(1, &_apResourceGrayPxielBO[nPyrLevel_], 0));
size_t nSize;
cudaSafeCall( cudaGraphicsResourceGetMappedPointer((void **)&pDev, &nSize , _apResourceGrayPxielBO[nPyrLevel_]));
cv::cuda::GpuMat cvgmRGBA( cvgmRGB_.size(), CV_8UC1, pDev);
cvgmRGB_.copyTo(cvgmRGBA);
cudaSafeCall( cudaGraphicsUnmapResources(1, &_apResourceGrayPxielBO[nPyrLevel_], 0) );
//texture mapping
glBindTexture(GL_TEXTURE_2D, uTexture);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, _auGrayPixelBO[nPyrLevel_]);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RED, cvgmRGB_.cols, cvgmRGB_.rows, 0, GL_RED, GL_UNSIGNED_BYTE, NULL);
errorDetectorGL();
glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
glBindTexture(GL_TEXTURE_2D, 0);
}
return uTexture;
}//gpuMapRgb2PixelBufferObj
std::vector<bbox_t> tracking_flow(cv::Mat dst_mat, bool check_error = true)
{
if (sync_PyrLKOpticalFlow_gpu.empty()) {
std::cout << "sync_PyrLKOpticalFlow_gpu isn't initialized \n";
return cur_bbox_vec;
}
int const old_gpu_id = cv::cuda::getDevice();
if(old_gpu_id != gpu_id)
cv::cuda::setDevice(gpu_id);
if (dst_mat_gpu.cols == 0) {
dst_mat_gpu = cv::cuda::GpuMat(dst_mat.size(), dst_mat.type());
dst_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1);
}
//dst_grey_gpu.upload(dst_mat, stream); // use BGR
dst_mat_gpu.upload(dst_mat, stream);
cv::cuda::cvtColor(dst_mat_gpu, dst_grey_gpu, CV_BGR2GRAY, 1, stream);
if (src_grey_gpu.rows != dst_grey_gpu.rows || src_grey_gpu.cols != dst_grey_gpu.cols) {
stream.waitForCompletion();
src_grey_gpu = dst_grey_gpu.clone();
cv::cuda::setDevice(old_gpu_id);
return cur_bbox_vec;
}
////sync_PyrLKOpticalFlow_gpu.sparse(src_grey_gpu, dst_grey_gpu, prev_pts_flow_gpu, cur_pts_flow_gpu, status_gpu, &err_gpu); // OpenCV 2.4.x
sync_PyrLKOpticalFlow_gpu->calc(src_grey_gpu, dst_grey_gpu, prev_pts_flow_gpu, cur_pts_flow_gpu, status_gpu, err_gpu, stream); // OpenCV 3.x
cv::Mat cur_pts_flow_cpu;
cur_pts_flow_gpu.download(cur_pts_flow_cpu, stream);
dst_grey_gpu.copyTo(src_grey_gpu, stream);
cv::Mat err_cpu, status_cpu;
err_gpu.download(err_cpu, stream);
status_gpu.download(status_cpu, stream);
stream.waitForCompletion();
std::vector<bbox_t> result_bbox_vec;
if (err_cpu.cols == cur_bbox_vec.size() && status_cpu.cols == cur_bbox_vec.size())
{
for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
{
cv::Point2f cur_key_pt = cur_pts_flow_cpu.at<cv::Point2f>(0, i);
cv::Point2f prev_key_pt = prev_pts_flow_cpu.at<cv::Point2f>(0, i);
float moved_x = cur_key_pt.x - prev_key_pt.x;
float moved_y = cur_key_pt.y - prev_key_pt.y;
if (abs(moved_x) < 100 && abs(moved_y) < 100 && good_bbox_vec_flags[i])
if (err_cpu.at<float>(0, i) < flow_error && status_cpu.at<unsigned char>(0, i) != 0 &&
((float)cur_bbox_vec[i].x + moved_x) > 0 && ((float)cur_bbox_vec[i].y + moved_y) > 0)
{
cur_bbox_vec[i].x += moved_x + 0.5;
cur_bbox_vec[i].y += moved_y + 0.5;
result_bbox_vec.push_back(cur_bbox_vec[i]);
}
else good_bbox_vec_flags[i] = false;
else good_bbox_vec_flags[i] = false;
//if(!check_error && !good_bbox_vec_flags[i]) result_bbox_vec.push_back(cur_bbox_vec[i]);
}
}
cur_pts_flow_gpu.swap(prev_pts_flow_gpu);
cur_pts_flow_cpu.copyTo(prev_pts_flow_cpu);
if (old_gpu_id != gpu_id)
cv::cuda::setDevice(old_gpu_id);
return result_bbox_vec;
}
