Real-time head pose estimation for driver assistance system using low-cost on-board computer

We propose a fast head pose estimation method using monocular video. It is highly optimized for on-board computers and for driving situations, which is applicable to existing low-cost on-board computer for cars and suitable for high real-time driver assistance systems. In our algorithm pipeline, the face detection step eliminates slow floating point computations using pixel intensity binary test, and reduce search scope effectively. In the face alignment step, we utilize the high performance of local binary feature and extend the single pose regression model to handle large rotations. The pose estimation step uses a mean rigid face model to calculate head pose fast by solving 2D-3D correspondence. To reduce computation further, we bypass or simplify pipeline steps using previous frame result, and redo the full pipeline adaptively. Experiments show our method is more efficient than existing approaches, which makes high real-time applications for on-board computers possible.