Research > Tracking |
Dynamic Programming Tracking
This tracking algorithm prevents taking (possibly wrong) local decisions, because the tracking is done at the end of a sequence by making a traceback of the decisions to reconstruct the best path t (x, y). This can be compared to time alignment in speech recognition.
Published paper related to this work:
-
P. Dreuw, J. Forster, and H. Ney. Tracking Benchmark Databases for Video-Based Sign Language Recognition. In ECCV International Workshop on Sign, Gesture, and Activity (SGA), pages 286-297, Crete, Greece, September 2010.
-
P. Dreuw, J. Forster, T. Deselaers, and H. Ney. Efficient Approximations to Model-based Joint Tracking and Recognition of Continuous Sign Language. In IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-6, Amsterdam, The Netherlands, September 2008.
© -
T. Deselaers, P. Dreuw, and H. Ney. Pan, Zoom, Scan -- Time-coherent, Trained Automatic Video Cropping. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1-8, Anchorage, AK, USA, June 2008.
-
P. Dreuw, T. Deselaers, D. Rybach, D. Keysers, and H. Ney. Tracking Using Dynamic Programming for Appearance-Based Sign Language Recognition. In IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 293-298, Southampton, UK, April 2006.
©
Dynamic programming tracking performs well when one wants to track an object with many occlusions, information gaps or for offline tracking. It may also be used with non- static background or multiple target objects in the foreground.
Dynamic programming tracking on large images is very time consuming, when making a full search of all possible tracking rectangles. In a first step we developed a simple tracking algorithm with fixed tracking rectangle size.
Since the dynamic programming tracking is an algorithm that can use any image score function, the minimum search window size I x J must be greater than one in order to calculate a score of at least one pixel, and in order to center the window, it should be of odd size. Thus for discrete functions or distributions, the minimum window size is set at 3x3.
In each dynamic programming algorithm, we need a recursion equation as in an HMM to calculate the best score over the whole image sequence, and to reconstruct the path which achieved this score.
An example where simply motion is used to track a hand signing some Palm's Graffiti didgits can be found in the follwing video clip. We used a partial traceback in order to reduce the computation power and memory requirements. The other examples are extracted from the RWTH-BOSTON-104 sign language database.
Philippe Dreuw Last modified: 2011-04-07 10:30:07 . Disclaimer. Created Wed Dec 22 18:04:32 CET 2004