http://feelmare.blogspot.kr/2014/04/opencv-study-calcopticalflowfarneback.html
In the GPU mode, the return value of the function is two vector direction, that is x direction and y direction.
In the GPU mode, functions of resize, cvtColor cannot copy to same variable of gpumat.
And gputmat cannot access to the pixel point using at
GPU is more fast about 10 times than cpu mode in my computer environment.
refer to this example source code and video.
#include < stdio.h>
#include < iostream>
#include < opencv2\opencv.hpp>
#include < opencv2/core/core.hpp>
#include < opencv2/highgui/highgui.hpp>
#include < opencv2/video/background_segm.hpp>
#include < opencv2\gpu\gpu.hpp>
#ifdef _DEBUG
#pragma comment(lib, "opencv_core247d.lib")
#pragma comment(lib, "opencv_imgproc247d.lib") //MAT processing
#pragma comment(lib, "opencv_objdetect247d.lib") //HOGDescriptor
#pragma comment(lib, "opencv_gpu247d.lib")
//#pragma comment(lib, "opencv_features2d247d.lib")
#pragma comment(lib, "opencv_highgui247d.lib")
#pragma comment(lib, "opencv_ml247d.lib")
//#pragma comment(lib, "opencv_stitching247d.lib");
//#pragma comment(lib, "opencv_nonfree247d.lib");
#pragma comment(lib, "opencv_video247d.lib")
#else
#pragma comment(lib, "opencv_core247.lib")
#pragma comment(lib, "opencv_imgproc247.lib")
#pragma comment(lib, "opencv_objdetect247.lib")
#pragma comment(lib, "opencv_gpu247.lib")
//#pragma comment(lib, "opencv_features2d247.lib")
#pragma comment(lib, "opencv_highgui247.lib")
#pragma comment(lib, "opencv_ml247.lib")
//#pragma comment(lib, "opencv_stitching247.lib");
//#pragma comment(lib, "opencv_nonfree247.lib");
#pragma comment(lib, "opencv_video247d.lib")
#endif
using namespace cv;
using namespace std;
void drawOptFlowMap_gpu (const Mat& flow_x, const Mat& flow_y, Mat& cflowmap, int step, const Scalar& color) {
for(int y = 0; y < cflowmap.rows; y += step)
for(int x = 0; x < cflowmap.cols; x += step)
{
Point2f fxy;
fxy.x = cvRound( flow_x.at< float >(y, x) + x );
fxy.y = cvRound( flow_y.at< float >(y, x) + y );
line(cflowmap, Point(x,y), Point(fxy.x, fxy.y), color);
circle(cflowmap, Point(fxy.x, fxy.y), 1, color, -1);
}
}
int main()
{
//resize scale
int s=4;
unsigned long AAtime=0, BBtime=0;
//variables
Mat GetImg, flow_x, flow_y, next, prvs;
//gpu variable
gpu::GpuMat prvs_gpu, next_gpu, flow_x_gpu, flow_y_gpu;
gpu::GpuMat prvs_gpu_o, next_gpu_o;
gpu::GpuMat prvs_gpu_c, next_gpu_c;
//file name
char fileName[100] = ".\\mm2.avi"; //Gate1_175_p1.avi"; //video\\mm2.avi"; //mm2.avi"; //cctv 2.mov"; //mm2.avi"; //";//_p1.avi";
//video file open
VideoCapture stream1(fileName); //0 is the id of video device.0 if you have only one camera
if(!(stream1.read(GetImg))) //get one frame form video
return 0;
//////////////////////////////////////////////////////////////////////////////////////////////
//resize(GetImg, prvs, Size(GetImg.size().width/s, GetImg.size().height/s) );
//cvtColor(prvs, prvs, CV_BGR2GRAY);
//prvs_gpu.upload(prvs);
//////////////////////////////////////////////////////////////////////////////////////////////
//gpu upload, resize, color convert
prvs_gpu_o.upload(GetImg);
gpu::resize(prvs_gpu_o, prvs_gpu_c, Size(GetImg.size().width/s, GetImg.size().height/s) );
gpu::cvtColor(prvs_gpu_c, prvs_gpu, CV_BGR2GRAY);
/////////////////////////////////////////////////////////////////////////////////////////////
//dense optical flow
gpu::FarnebackOpticalFlow fbOF;
//unconditional loop
while (true) {
if(!(stream1.read(GetImg))) //get one frame form video
break;
///////////////////////////////////////////////////////////////////
//resize(GetImg, next, Size(GetImg.size().width/s, GetImg.size().height/s) );
//cvtColor(next, next, CV_BGR2GRAY);
//next_gpu.upload(next);
///////////////////////////////////////////////////////////////////
//gpu upload, resize, color convert
next_gpu_o.upload(GetImg);
gpu::resize(next_gpu_o, next_gpu_c, Size(GetImg.size().width/s, GetImg.size().height/s) );
gpu::cvtColor(next_gpu_c, next_gpu, CV_BGR2GRAY);
///////////////////////////////////////////////////////////////////
AAtime = getTickCount();
//dense optical flow
fbOF.operator()(prvs_gpu, next_gpu, flow_x_gpu, flow_y_gpu);
//fbOF(prvs_gpu, next_gpu, flow_x_gpu, flow_y_gpu);
BBtime = getTickCount();
float pt = (BBtime - AAtime)/getTickFrequency();
float fpt = 1/pt;
printf("%.2lf / %.2lf \n", pt, fpt );
//copy for vector flow drawing
Mat cflow;
resize(GetImg, cflow, Size(GetImg.size().width/s, GetImg.size().height/s) );
flow_x_gpu.download( flow_x );
flow_y_gpu.download( flow_y );
drawOptFlowMap_gpu(flow_x, flow_y, cflow, 10 , CV_RGB(0, 255, 0));
imshow("OpticalFlowFarneback", cflow);
///////////////////////////////////////////////////////////////////
//Display gpumat
next_gpu.download( next );
prvs_gpu.download( prvs );
imshow("next", next );
imshow("prvs", prvs );
//prvs mat update
prvs_gpu = next_gpu.clone();
if (waitKey(5) >= 0)
break;
}
}
Hello,
ReplyDeleteThank you very much for sharing your job.
I would like to know which pc did you need to obtain this result ?
(RAM, processor,graphics board)
Thank you very much !
Camille
Thank you for visiting my blog.
ReplyDeleteMy computer spec is
xeon cpu e5-2609
ram 16gb
quadro k2000
but
i7, 8ram, geforce graphic spec is enough to obtain that performance.
Hi there! nice setup. You know, i'm probably gonna have to work with optical flow and cnn really soon, and I have access to a server with 6 core xeon cpu, and a quadro k4200.
ReplyDeleteI'd like to know, did you do any benchmarks? how many fps did you got?
Nice blog btw :)