Head Pose Estimation from Single Camera

Alessandro Marianantoni


Abstract

The study is focused on the recovery of full-motion (3 rotations and 3 translations) of the head using a single camera.The method uses a Cylindrical model to approximate the head and the motion is obtained by optical flow using the Lucas-Kanade technique. To deal with non-rigid motion and occlusion it is implemented as Iteratively Re-weighted Least Squares in conjuction with an image gradient. To reduce the aperture problem I am using a regularization-based approach. I am also using background subtraction by a statistical model. The traking program was evaluated using a webcamera on the monitor.


Maintained by Philippos Mordohai