Fast Head Pose Estimation via Rotation-Adaptive Facial Landmark Detection for Video Edge Computation

The human head pose estimation is an important and challenging problem, which provides the estimation of the head posture in 3D space from 2D image. It is a crucial technique for face recognition, gaze estimation, facial attribute recognition, etc. However, fast head pose estimation executing on the terminal for video edge computation has many challenges due to the computational complexity of the existing algorithms. In this paper, we propose a fast head pose estimation method based on a novel Rotation-Adaptive facial landmark detection powered by Local Binary Feature (RALBF). The landmark detection method is structured through fusing the prior of the rotation information provided by the Progressive Calibration Networks (PCN) face detector to a Local Binary Feature (LBF) based landmark detection method, which improves the robustness against head pose variations and simultaneously keep the computing efficiency. RALBF is trained and tested on 300W dataset and AFLW2000 dataset, it is verified by the accuracy evaluation that RALBF performs better than LBF. To improve the speed of head pose estimation, the 68, 51 and 10 landmarks distribution schemes are explored and compared on speed and accuracy. In the 10 landmarks scheme, the head pose estimation running once only takes 8.3ms on Intel i7-6700HQ CPU and takes 21.8ms on HiSilicon SoC Hi3519AV100, and the average error of Euler angle is 5.9973° when the face yaw angle is between ±35° on AFLW2000 3D dataset. Experiments demonstrate our approach performing well on real scenes.