Fasthpe : a recipe for quick head pose estimation

Abstract

Estimating the head orientation of a person from a single camera is an important step for human-computer interaction, especially for widely available laptops and hand-held devices. This work aims to track a human face and estimate its orientation in the 6 degrees of freedom from an uncalibrated monocular camera, keeping the user free of any devices or wires. We propose a novel algorithm based on existing computer vision techniques for a real-time (2ms) head pose estimation system, which can start and recover from failure automatically without any previous knowledge of the user’s appearance or location. We demonstrate that this computationally efficient pose estimation system is able to track the continuous roll, yaw, and pitch angles within absolute errors of 3.03, 5.27 and 3.91 degrees respectively. We show that with the tracking of only four face features, it is possible to obtain continuous head orientation measurements in real-time (2ms).