Multi-stage Real-time Human Head Pose Estimation

Abstract

The paper explores a human head pose estimation method combining face detection and face dichotomy optimization algorithm. The estimation of head pose is mainly to obtain the angle information of face orientation. The human head pose estimation algorithm in this paper mainly estimates the 3D Euler Angle of the input face patch. That is to say, a regressor is trained in a data-driven way, which can directly predict the input face blocks. In this paper, we use three models to make the final prediction of human head posture. First, the first model is used to detect faces. So as to prevent the loss of face detection in realtime face detection, we set a small threshold on the face detection model. Second, on the basis of face detection, a face dichotomy model is added to optimize the detected face to determine whether it is a face. Third, if the result of the second step is a face, the head pose estimation of the detected face is carried out. This paper mainly aims at the change of head pose angle to judge people's attention, and puts forward multi-stage head pose estimation to meet the demand of real-time engineering calculation. In this paper, the first novelty is that the average speed on Windows with CPU(i5-7500 3.40GHz) is 35ms/frame, and the second novelty is that we have collected 64101 face images under the surveillance camera, and the third novelty is that we have mixed depthwise separable convolution for head pose recognition. The method is tested on the UMDFaces Dataset, and the consequences show that compared with extra approaches, the proposed method can obtain improvements in efficiency.