An Alternative Approach to Recovering 3D Pose Information from 2D Data

A novel approach is presented for real-time pose estimation of a known object being observed through a single camera. The future aim is to develop the method outlined here into a robust closed-form solution suitable for implementation on a monocular face tracking system. However, the method outlined in this paper offers a generic 2D to 3D pose estimation technique which is not limited to face tracking applications. The approach requires the ob served object to consist of at least six 3D points. It uses true perspective projection and provides a means towards a direct solution, without the need for iteration. While the limited information available from a monocular vision system suggests that stereo vision is more suitable for pose recognition problems these stereo camera must still be calibrated. Camera calibration is an equivalent problem to monocular pose estimation, and must be performed on each camera individually, and so the approach discussed will also be useful in the calibration of a stereo system.