Stereo based 3d head pose tracking using the scale invariant feature transform

Official URL: http://risc01.sabanciuniv.edu/record=b1221513 (Table of Contents)

In this thesis a new stereo-based 3D head tracking technique, based on scale invariant feature transform (SIFT) features, that is robust to illumination changes is proposed. Also two major tracking techniques, one based on normal flow constraints (NFC) and a 3D registration-based method, based on iterative closest point (ICP) algorithm, are reviewed and compared against the proposed technique. A 3D head tracker is very important for many vision applications. The resulting tracker output parameters can be used to generate a stabilized view of the face that can be used as input to many existing 2D techniques such as facial expression analysis, lip reading, eye tracking, and face recognition. Our system can automatically initialize using a simple 2D face detector. We extract salient points from the intensity images using SIFT features and match them between frames. Together with the depth image and the matched features we obtain 3D correspondences. Using the unit quaternion method, we recover the 3D motion parameters. Our proposed method outperforms both NFC and ICP on translations; and performs as good as NFC on rotations. Experimentally, the proposed system is less likely to drift than NFC and ICP over long sequences and is robust to illumination changes. We present experiments to test the accuracy of our SIFT-based 3D tracker on sequences of synthetic and real stereo images.