Temporal Head Pose Estimation From Point Cloud in Naturalistic Driving Conditions

Head pose estimation is an important problem as it facilitates tasks such gaze and attention modeling. In the automotive context, head provides crucial information about driver’s mental state, including drowsiness, distraction attention. It can also be used for interaction with in-vehicle infotainment systems. While computer vision algorithms using RGB cameras are reliable in controlled environments, a challenging car due to sudden illumination changes, occlusions large rotations that common vehicle. These issues partially alleviated by depth cameras. rotation trajectories continuous temporal dependencies. Our study leverages this observation, proposing novel deep learning model from point cloud. The approach extracts discriminative feature representation directly cloud data, leveraging 3D spatial structure of face. frame-based representations then combined bidirectional long short term memory (BLSTM) layers. We train on newly collected xmlns:xlink="http://www.w3.org/1999/xlink">multimodal driver monitoring (MDM) dataset, achieving better results compared non-temporal state-of-the-art models images. further show quantitatively qualitatively incorporating improvements not only accuracy, but smoothness predictions.