Estimation of Driver Head Yaw Angle Using a Generic Geometric Model

Head yaw angle is an important indicator of a driver's distraction. Forward collision warning systems incorporate head yaw angle estimators for detecting driver's head-offroad glances. This paper presents a generic geometric model for head yaw estimation. The generic model is customized into 12 different models. The performance of the cylindrical and ellipsoidal models is improved by introducing nose projection and truncation. The yaw estimation problem is formulated using a single variable. This formulation provides better visualization of the cylindrical and ellipsoidal models. It is shown that the ellipsoidal model in literature is indeed a cylindrical model with nose projection. We have also proved that, under the ellipsoidal/cylindrical framework, the estimated yaw angle is independent of shifting the center of rotation along depth. Four previous works are shown as the customization of the proposed generic model. The performance of the proposed models are evaluated using four standard head pose datasets and an ADAS dataset. The proposed nose-projected truncated ellipsoidal model outperforms the state-of-the-art geometric models. Further, we have studied the impact of model parameters and inaccurate facial feature localization through simulation experiments.