Homography-based face orientation determination from a fixed monocular camera

Abstract

This paper presents a framework for determining the orientation of human faces with a fixed monocular camera which can be used for the purposes of the gaze tracking afterwards. We use homography relation between two views/frames to handle with the lack of depth information. In order to compensate for the lack of depth information in the relationships between the 2D images in the image plane and the 3D Euclidean space, we present a complete vision-based approach to pose estimation. The homography relates corresponding points captured at two different locations of the face and determines the relationships between the two locations using pixel information and intrinsic parameters of the camera. In order to determine the mapping between the two images, it is assumed that in each frame in the video sequence, we are able to locate, extract and labeled four feature points of the face located at virtual plane attached to the face. Face detection and facial feature extraction are executed with Viola-Jones method. The verification stage for face detection use combined cascade of neural network classifiers uses the convolutional neural network.