Head Pose Estimation via mmWave Radar

Head pose estimation in autonomous driving enhances safety by monitoring real-time driver attention and behavior. However, mainstream vision-based methods encounter challenges like susceptibility to spoofing, lighting sensitivity, and privacy risks. Wearable solutions also face comfort and fitting issues. In this study, we introduce mmHead, a novel human head posture estimation system using mmWave radar technology. In the initial phase, we utilized FFT and static interference removal techniques to extract head pose features, aiming to mitigate environmental noise and user variations. These features were then inputted into a Convolutional Neural Network (CNN) for head pose estimation, where convolutions were performed along length, width, and depth dimensions. Using the TI 6843AOP mmWave radar, we collected a dataset of 70,000 frames from various subjects. Evaluation by multiple participants showed mmHead achieving an angular deviation within 6° for head pose estimation across all three angles.