Research Of Video-based Mood Detection System

Facial expressions play a very important role in conveying information as one of the non-verbal behavior. In the past two decades automatic facial expression recognition has been the hot topics of human-computer interaction. Among Facial expressions, smile is the most representative and most special. Laughter can be seen as a hint of good mood. Therefore, this paper will mainly focus on smile detection which can be seen as mood detection, rather than developing a comprehensive expression recognizer. Smile Detection has enormous potential applications. For example, record the person who is watching a commercial advertisement by camera and automatically analyze whether he or she is smiling or we can say in good mood. It can be seen that whether he or she enjoy the advertisement.1. This paper firstly confirms the three steps of the framework of the mood detection system. They are face detection and location, expression feature extraction and expression classification. It is confirmed that the face detection and location has been very mature.The face detection method based on Haar features and the Adaboost algorithm proposed by Viola et al. is studied here and a face detector is trained according to this method.2. After a survey of the process of face detection it can be seen that the detection process can be improved in accordance with the characteristics of the video. In the video recorded by a fixed camera which is installed at an inclined angle the size of human face is basically identical at the same horizontal line. And the size of human face will be bigger in the video frame and human face will be closer to the bottom of video frame when the human is walking close to the camera than away from it. According to this a video-based method of setting the searching window of human face is proposed. Video test indicates that the detection speed has been greatly enhanced meanwhile the number of false detections are reduced. 3. After a survey of color space and color model, it is found that the further validation of the areas judged as faces by the face detector can be done by the information of YCbCr color space and skin color range model. Some false alarms can be excluded by skin color information.4. Using Haar features and Gentle Adaboost algorithm to extract mood characteristics and classify them, the final mood detection system reachs the speed of15frames/second when the size of video is320x240. It can meet the basic needs of real-time processing.