Abstract:
We describe a novel corner tracking system which substantially reduces algorithmic errors, while still maintaining a low dropout rate. This is done by combining the results of a conventional correspondence-based corner tracker with a novel relaxation-based tracker. Machine learning techniques are applied to further improve performance of the individual trackers. Preliminary results with synthetic data indicate the system has half the dropout rate of a correspondence based system.
 

Downloads:

• A gzipped postscript version of the paper is available from here. 
• An HTML overview of this research is available here.