Christopher Evans. OpenSURF. http://code.google.com/p/opensurf1/

http://sites.google.com/site/chrisevansdev/files/opensurf.pdf

Refering to the paper I linked:

Our library lacks of the Taylor refinement described in section 1.2.3 (Accurate Interest Point Localisation) and of the orientation assigment for the descriptor (section 1.3.1, Orientation Assignment)

The project has a complete structure refering to the interfaces and it is easy extendable by implementing different classes from the provided interfaces. In particular, you just have to write a new realization of the Descriptor and Detector interfaces.

As far as the video analysis, we used a producer-consumer pattern (using separating threads), with the producer providing frames from the video and the consumer analysing them with the P-SURF algorithm.

We used this library to work with videos:

JMCvideo: http://www.mat.ucsb.edu/~a.forbes/PROCESSING/jmcvideo/jmcvideo.html

plus the standard one provided by Processing.

Every single frame is analyzied by P-SURF. We compare the resulting interesting points to our reference interest points obtained from the target image (the banner we wanted to find in images) by calculating the euclidean norm of the distance between the features with a certain tolerance. If there were more than a certain threshold of matching ( I think we used like more than 10% ), we said that the banner was found.

If you have other question, me and my friend Claudio Fantacci will be happy to help you.

Best regards

1. private double euclideanDistance(   Interest_Point target,

               Interest_Point another_point) {

   float[] targetDescriptor = target.getDescriptorOfTheInterestPoint();

   float[] theOtherOneDescriptor = another_point

   .getDescriptorOfTheInterestPoint();

   
   

   double distance = 0;

   
   

   for (int i = 0; i < 64; i++) {

   
   

   distance += Math.pow(

   targetDescriptor[i] - theOtherOneDescriptor[i], 2);

   
   

   }

   
   

   return Math.sqrt(distance);

   }

1. 
   

   private ArrayList<Interest_Point> imageToSearchInto;

   private ArrayList<Interest_Point> matchingInterestPoints = new ArrayList<Interest_Point>();

   //Array cointaining all the IP of the target image

   private ArrayList<Interest_Point> InterestPointOfTheBanner;

   
   

   public void findMeHere() {

   
   

   Interest_Point target;

   
   

   for (int i = 0; i < InterestPointOfTheBanner.size(); i++) {

   
   

   target =  InterestPointOfTheBanner.get(i);

   
   

                           //Look for the best match in imageToSeachInto

   Interest_Point best = getBest(target, imageToSearchInto);

           if (euclideanDistance(target, best) < 0.7) {

   
   

                                           // Found! Adding to matching IP

   matchingInterestPoints.add(best);

                                           // I remove the IP to the ones to be used for new matching

   imageToSearchInto.remove(best);

   
   

   }

   }

   
   

   }

                   // 10% of threshold

   if ( InterestPointOfTheBanner.size()/matchingInterestPoints.size() > 0.1){

   
   

                           // Target image found!

   
   

                   }

   
   

   }