A novel regularization method for optical flow-based head pose estimation
Sebastian Vater, Guillermo Mann, Fernando Puente León · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2015
This paper presents a method for appearance-based 3D head pose tracking utilizing optical flow computation. The task is to recover the head pose parameters for extreme head pose angles based on 2D images. A novel method is presented that enables a robust recovery of the full motion by employing a motion-dependent regulatory term within the optical flow algorithm. Thereby, the rigid motion parameters are coupled directly with a regulatory term in the image alignment method affecting translation and rotation independently. The ill-conditioned, nonlinear optimization problem is stabilized by the proposed regulatory term yielding suitable conditioning of the Hessian matrix. It is shown that the regularization corresponding to the motion parameters can be extended to full 3D motion consisting of six parameters. Experiments on the Boston University head pose dataset demonstrate the enhancement of robustness in head pose estimation compared to conventional regularization methods. Using well-defined values for the regulatory parameters, the proposed method shows significant improvement in headtracking scenarios in terms of accuracy compared to existing methods.