Abstract

Since fatigue and drowsy can result in severe disaster, early detection is of great importance. Among the extractable features from driver such as facial expression and eye blink rate, head pose estimation has been widely used as a key element for driver monitoring due to its robustness to interpersonal variation. Most of the research about head pose estimation has been conducted with 2D RGB images based on facial key-points detection. However, the methods based on 2D RGB images are vulnerable to occlusion and illumination changes. To alleviate these problems, head pose estimation method using point cloud is proposed with deep learning to infer 6D vector. In the experiment, it is shown that the performance of our method exceeds that of RGB-based methods by +5.44 in terms of MAE on BIWI dataset.