3D Face Tracking Using Stereo Cameras with Whole Body View

Abstract

All visual tracking tasks associated with people tracking are in a great demand for modern applications dedicated to make human life easier and safer. In this thesis, a special case of people tracking - 3D face tracking in whole body view video is explored. Whole body view video means that the tracked face typically occupies not more than 5-10% of the frame area. Currently there is no reliable tracker that can track a face in long-term whole body view videos with luminance cameras in the 3D space.

I followed a non-classical approach to designing a 3D tracker: first a 2D face tracking algorithm was developed in one view and then extended into stereo tracking. I recorded and annotated my own extensive dataset specifically for 2D face tracking in whole body view video and evaluated 17 state of the art 2D tracking algorithms. Based on the TLD tracker, I developed a face adapted median flow tracker that shows superior results compared to state of the art generic trackers. I explored different ways of extending 2D tracking into 3D and developed a method of using the epipolar constraint to check consistency of 3D tracking results. This method allows to detect tracking failures early and improves overall 3D tracking accuracy. I demonstrated how a Kinect based method can be compared to visual tracking methods and compared four different visual tracking methods running on low resolution fisheye stereo video and the Kinect face tracking application.

My main contributions are: - I developed a face adaptation of generic trackers that improves tracking performance in long-term whole body view videos. - I designed a method of using the epipolar constraint to check consistency of 3D tracking results.