roboforum.ru

Код: Выделить всё • Развернуть

#include <cv.h>
#include <cxcore.h>
#include <highgui.h>
#include <time.h>

#pragma comment(lib,"highgui.lib" )
#pragma comment(lib,"cv.lib" )
#pragma comment(lib,"cxcore.lib" )

void main (void)
{
   IplImage *src, *tmp;
   CvSeq* contour = 0; 

   cvNamedWindow( "test",1 );
   cvNamedWindow( "test1",1 );
   cvNamedWindow( "test2",1 );

   CvCapture *cap = cvCaptureFromCAM(0);
   for(;;){
      tmp = cvQueryFrame(cap);
      cvShowImage( "test1",tmp );

                        // у меня 640х480 если у вас меньше то следующий 6 строк заремарьте и все будет работать идеально. Сдесь я делаю сглаживание
                        // потому что мая камера очень далека от идеала и шумы не позволяют чтото видеть
      IplImage *pyr = cvCreateImage( cvSize(320,240),8,3 );
      cvPyrDown( tmp, pyr, 7 );
      cvPyrUp( pyr, tmp, 7 );
      cvReleaseImage( &pyr );
      cvShowImage( "test2",tmp );

      src = cvCreateImage( cvGetSize(tmp),tmp->depth,1 );
      int COI = cvGetImageCOI( tmp );
      cvSetImageCOI( tmp,2 );
      cvCopy( tmp,src,0 );
      cvSetImageCOI( tmp,COI );

      cvCanny( src,src,30,70,3 );
      cvDilate( src,src,0,1 );
      cvErode( src,src,0,1 );
      
      cvZero( tmp );
      CvMemStorage* contour_storage = cvCreateMemStorage(0);

      CvContourScanner blobs = cvStartFindContours( src, contour_storage, sizeof(CvContour), CV_RETR_CCOMP,CV_CHAIN_APPROX_NONE );
      while( (contour=cvFindNextContour(blobs)) != NULL ){
         CvScalar outer = CV_RGB( 0, 255, 255 );
         CvScalar inner = CV_RGB( 255, 255, 255 );
         cvDrawContours( tmp, contour, outer, inner, 0, 0 );
      }
      contour = cvEndFindContours( &blobs );
   
      cvReleaseMemStorage( &contour_storage );

      cvShowImage( "test",tmp );
      cvWaitKey(1);

      cvReleaseImage( &src );
   }

   /* Releases contour scanner and returns pointer to the first outer
   contour */
   cvReleaseCapture( &cap );

   return;
}

Кстати, в OpenCV по моему 3-4 метода реализовано частично из семейства optical-flow, правда вроде без калибровки.

Actually, for one of our current projects we are now studying many
different optical flow algorithms, such as ones by Bruhn & Weickert
(http://www.mia.uni-saarland.de/Publications/bruhn-iccv05.pdf), or
Gunnar Farneback (http://www.isy.liu.se/~gf/). In particular, in this
paper by G. Farneback
http://www.cvl.isy.liu.se/ScOut/Publications/Papers/iccv01_gf.pdf
he also uses segmentation, but this is not color segmentation, this is
motion segmentation, it is built simultaneously with motion estimation,
that is, the final regions correspond to the regions that move as a whole.

We plan to include some of these implementations into the library pretty
soon, old algorithms (HS, LK, BM) will be thrown away. And to make sure
that we put the best algorithms, we plan to do some extensive
comparison/benchmarking.

Интересная инфа, а что за Proposed Method? Личная разработка, я так понимаю?

Anyway, it is very nice to see some alternative approaches.
The problem with segmentation-based algorithms is that they are not
quite robust to some noise or lighting changes - once you draw a
boundary between components, you threw away some information that may be
valuable. Also, there are some kinds of videos where segmentation does
hardly make sense, e.g. medical images, or videos from satellites.

On the other hand, there are some algorithms that try to do segmentation
and optical flow estimation simultaneously, that is, they segment image
into the regions using the motion consistency, rather than color, or
brightness, or texture or whatever.
And once you have such segmentation, you may compute the motion in each
region very accurately using least squares or some robust estimator.

Another possible direction of research is to take a look at 3d
algorithms, i.e. algorithms that consider not only 2 frames, the
previous and the next, but a few frames before and after the current
frame.

=DeaD= писал(а):

EDV писал(а):Может у них уже лучше есть и они его уже включают в OpenCV? Всё-таки люди тоже вроде не профаны...