Analysis of Facial Landmark Features to determine the best subset for finding Face Orientation

Abstract

The number of applications which use human face analysis are going up by the day and face orientation or pose detection is an important and upcoming research in this area. This paper uses a mathematical technique which compares real world coordinates of facial feature points with that of 2D points obtained from an image or live video using a projection matrix and Levenberg-Marquardt optimization to determine the Euler angles of the face. Further, this technique is used to find the best set of facial landmarks which give the maximum range of detection. The preliminary steps of the face orientation technique are face detection and facial landmark detection. For face detection, the Haar Cascade and Deep Neural Network techniques are experimented. Based on the analysis it is concluded that DNN is more robust, accurate and optimal. Facial landmarks are extracted by passing an image or video frame through a cascade of pre-trained regression trees. After analyzing various sets of facial features for their use in face orientation detection techniques and testing the results of each, a set of six facial points nose tip, chin, corner points of the eyes and corner points of the mouth are found to be enough for the algorithm to be able to detect the orientation of the face in a wide range of view with lesser computations.