Emotion Detection From Faces in Image Sequences Using Facial Point Tracking with Probabilistic Model Estimation with Particles

Abstract

Emotion detection is related to many fields such as behavioral study, rehabilitation, e-learning, etc. If computers can detect emotions, they will be able to play an important role in applications of these fields. Emotion detection process includes face detection, features extraction, and, emotion classification. Facial points are tracked using only the spacial information from the previous frame and texture information from the first frame. Texture information from the neutral face helps the tracking procedure to have accurate facial feature localization in each frame. The feature locations from the previous frame are used to predict the feature locations of the current frame. Location information from the previous frame helps the tracking procedure to track the facial feature locations of the current frame easily because the current feature locations tend to be located near the feature locations of the previous frame. Our expectation is to classify an emotion from the last frame (peak of emotion) in an image sequence. We use textures form the neutral face and the facial points from the previous frame to form a probabilistic model. After that, the facial points in each frame are assigned using the particle estimation to find the expected values of facial point locations. Then, we extract emotion from features produced by these points by using a classification. The benefit of using particles for probabilistic estimation is that finding expectation value of probabilistic model has low complexity.