•  Home 
• |
•  Over Gpsgek 
• |
•  Geocaching 
• |
•  Informatief 
• |
•  Links 
• |
•  Weblog 

Snelmenu

• Wat is Geocaching?
• Gevonden caches...
• Grafiekjes
• Stippenkaart
• Geplaatste caches...
• Eigen travelbugs en coins
• Geo d'Or-kaart
• Fietscaches
• Treincaches
• Afstandspictogrammen
• GCTool

Informatief

• Afstandspictogrammen
• Formule-service
• GCTool
• Geocachen en TomTom
• Kesj
• WGS84 <-> RD online
• WGS84 <-> RD script
• Feature tracking
• Interfacing a SD-card
• Visualizing with ParaView

Feitjes

• Gpsgek heeft al 1180 caches gevonden!
• Gpsgek heeft al 143 travelbugs gevonden!
• Gpsgek heeft al 2 travelertags gevonden!
• Er staan ruim 1365 foto's op deze site!

 

• Voeg toe aan favorieten

Feature Tracking and Handwritten Digit Recognition

This movie is about feature tracking and handwritten digit recognition. During the autumn of 2008, we combined two projects we did for two courses at Delft University of Technology. One was about Image Registration and Optical Flow for the course Advanced Digital Image Processing and one was about Handwritten Digit Recognition for the course Pattern Recognition.

Movie

The movie above shows a nice combination of the two projects, which was - off course - not required in the assignments but very funny to do. We developed our application in Matlab, which is a great tool to design such things. Yes, it is slow, but - as you see - with some optimizations fast enough to perform all calulations real-time.

Feature tracking

We implemented the feature tracking part using the Lucas-Kanade algorithm. Actually, we wrote a multi-level implementation to be able to track larger movements. The largest problem we encountered was to make the algorithm fast enough to be able to run real-time. This was caused by two things: 1) the large amount of interpolations slow down Matlab enormously; 2) the downsampling for the multi-level part took a lot of time. After applying some tricks to speed up the interpolation and downsampling, we got framerates of > 20 fps, which is more then fast enough...

Actually, I would like it to explain the whole implementation including explaining the source code, but I don't have time to do so. So, if - in the nearby future - I've some time left, I'll do, otherwise you can always contact me by email.

Some links about the Lucas-Kanade algorithm which are quite useful:

• B.D. Lucas and T. Kanade, An Iterative Image Registration Technique with an Application to Stereo Vision, Proceedings of Image Understanding Workshop, pp. 121-130 (1981).
  The original paper.
• S. Baker and I. Matthews, Lucas-Kanade 20 Years On: A Unifying Framework, Carnegie Mellon University.
  Okay, if you do not understand what the Lucas-Kanade algorithm does actually, you should read this great paper. Altough this implementation is not fast enough to perform real-time feature tracking, especially Fig. 2 make things very clear.
• J.Y. Bouguet, Pyramidal Implementation of the Lucas Kanade Feature Tracker, Intel Corporation.
  Gives a clear explanation how you could implement a multi-level approach.
• M.D. Steenwijk and H.A. van Zandbrink, Please find my head, Dec 2008.
  One of our intermediate results.
• Lucas-Kanade tracker
  This website shows a nice implementation and some cool movies.
   
• Matlab
  Especially the "Support"-section is very useful ;-).
• C.L.L. Hendriks, L.J. van Vliet, B. Rieger, M. van Genkel and R. Ligteringen, DIPimage and DIPlib, Delft University of Technology.
  The DIPimage-toolbox is very handy to speed up things!
• J. Lichtenauer, VfmDirect: Realtime Direct Show video capture function for Matlab, Delft University of Technology.
  This nice function helps to speed up the webcam capturing part.

Handwritten Digit Recognition

The pattern recognition part, so the part which recognizes the written digits, took net a lot more time to develop. We had to find out which combination of feature reduction method and classifier separates the digit classes the best. We tried a lot of combinations and found out that k-NN classification on a PCA feature reduced normalized pixel representation gives the lowest classification error (about 2.5%). This is short said, but the previous sentence contains in fact a few weeks calculation time by multiple desktops in order to show that this system is really the simplest and the best one (of which we tried, off course). Finally we have to thank our teachers for the development of the PRTools toolbox a lot, because this toolbox makes life a lot easier!

Some useful links on the pattern recognition part:

• Principal components analysis
  If you do not know what PCA is, this is quite useful to read.
• Y. LeCun and C. Cortes, The MNIST database of handwritten digits
  This is useful if you are looking for some training data.
   
• R.P.W. Duin, D. de Ridder, P. Juszczak, C. Lai, P. Paclick, E. Pekalska, D.M.J. Tax, PRTools: The Matlab Toolbox for Pattern Recognition, Delft University of Technology.
  If you're planning to do something with Pattern Recognition in Matlab, this toolbox is unmissable.
• D.M.J. Tax, DDtools, the Data Description Toolbox for Matlab (2008).
  And, if you are planning to do something with one-class classifiers, this toolbox adds a lot of functionality to PRTools.