Real-time Head Pose Estimation Based On Adaptive 3D Face Model

Head pose estimation, which is an important research direction in computer vision and determines the rotation angles involving three directions of face image in 3D space, has many applications including virtual reality, human-computer interaction, multi-view face recognition system, fatigue driving detection, and so forth.Nowadays, the methods of head pose estimation mainly include three categories:1) facial appearance-based; 2) classification-based; 3) models-based. In the early years, the work on head pose estimation mainly focused on facial appearance-based method. Recently, the classification-based and models-based models are widely focused. The former are usually used in static image which are hard to simulate the continuous process of head pose variation and hardly ensure the accuracy. Model-based methods are currently the most popular research methods and its main advantage is efficiency, and capable of high accuracy when facial landmarks is accurately localized.The related works of this thesis involve face detection, face alignment and head pose estimation around head pose estimation based on 3D face model. In face detection, face detection algorithm based on Haar features and AdaBoost cascade classifier is used to obtain face region firstly. In face alignment, the accuracy and efficiency of facial landmarks localization are considered and SDM (Supervised Descent Method) is applied to localize facial landmarks. In addition, we propose the facial landmarks tracking method. In head pose estimation, POSIT (Pose from Orthography and Scaling with Iterations) based on 3D face model is used because of the more flexible key points selection. We have considered the case of the poor localization of partial key points and propose an adaptive updated 3D face model method which applies POSIT to estimate head pose. In other words, the more accurate facial landmarks are selected adaptively and the 3D face model is updated according to the location of these facial landmarks.This thesis focuses on the research of head pose estimation methods based on template matching and based on POSIT, and the performance of POSIT and adaptive POSIT is compared. Moreover, we proposed the application of fatigue driving detection based on the SDM facial landmarks localization and head pose estimation, and used two groups of experiment to simulate its detection process. The experiments show that our methods can reach the speed of 15fps on Matlab2012a and 22fps on C++.