Pose estimation using cascade trees

Abstract

In this thesis, I implemented and extended a face detector, based on cascades of boosted features, for use in real time systems. The extensions are twofold. First, I designed a way of combining several cascades into a cascade tree, and showed how such a tree provides a powerful mechanism for combining detector efficiency and accuracy. When the training data has large variations, the cascade tree yields a faster detector, and when the data has only small variations, there is a distinct detection rate improvement. As a second extension, I designed a system for pose estimation based on an array of cascades. I performed an evaluation of this system and compared to normalized cross-correlation.