Selection of Reliable Features Subsets for Appearance-Based Tracking

Abstract

Efficient algorithms that track targets with a constant aspect (rigid objects, for example) are often based on appearance models. The simplest models linearly predict motion parameters from gray-scale variations measured at features. Choosing the features and training the predictor is done during a preliminary off-line stage. This paper presents several methods that improve the features selection process by filtering out some features from a given set. In particular, we are interested in the SVD-based subset selection procedure proposed by Golub and Van Loan. We show a significant improvement of tracking performance when our method filters Moravec, Harris, KLT or SUSAN features. We conclude that individually good selected features may not build a good subset and that a good spatial distribution of the features is paramount.