Real-time head pose estimation based on face geometry
Aditya Hosamani, Manoj Ravindra Phirke · 2020
This paper presents a novel head pose estimation technique based on face-geometry. The approach involves performing nose detection and threshold-based segmentation over near-infrared (NIR) input. NIR's inherent invariance to illumination changes makes it an ideal fit within automotive, avionics, mining, and many more such applications where the inconsistent lighting leaves visible light RGB cameras impractical. The pose angles are inferred from the shape of the segmented face and its nose location. Validation of the proposed approach is performed in a two-fold manner since annotated public NIR-based head-pose datasets are scarce. Firstly, the pure yaw and pitch angles are justified using the UPNA RGB head pose dataset. Secondly, an in-house captured and annotated NIR dataset is used for corroborating the pure roll angle. Comparison of the proposed approach with the two more commonly used head pose estimation algorithms viz., DLIB and OpenVINO over the NIR dataset reveal that the proposed approach outperforms the latter two with better accuracy, maximum range coverage, and lesser computation time, thereby, making it a suitable choice in applications such as driver monitoring and surveillance systems, especially during the night/low-light scenarios.