Real-time 3-D head motion estimation in facial image coding

Abstract

A simple procedure that uses only three feature points to infer 3D head motion from consecutive video frames is presented. In this procedure, a feature triangle formed using the three feature points is automatically calibrated, and an iterative method that simulates steepest descent method is applied to estimate one's head motion. A prediction algorithm is adopted in occasional cases where the estimated result is not acceptable. This procedure has been applied to live video with an update rate of 7 frames/sec (250 frames/sec without feature extraction) on a Pentium-II 233 MHz PC, independent of cameras and users.