An Efficient 3d Head Pose Inference from Videos

摘要

In this article, we propose an approach to infer the 3d head pose from a monocular video sequence. First, we employ a Gabor-Phase based displacement estimation technique to track face features (two inner eye corners, two wings, tip and root of the nose). The proposed method is based on the iterative Lowe's pose estimation technique using the six tracked image facial points and their corresponding absolute location in a 3d face model. As any iterative technique, the estimation process needs a good initial approximate solution that is found from orthography and scaling. With this method, the pose parameters are accurately obtained as continuous angular measurements rather than expressed in a few discrete orientations. Experimental results showed that under the assumption of a reasonable accuracy of facial features location, the method yields very satisfactory results.