Multi-view Face Detection And Tracking In Video Sequences

Human face detection and tracking is playing the basic role in video surveillance, human-computer interaction, and expression analysis fields. However, as the development of requirement in video surveillance and human-computer interaction fields, these fields requires the improvement of face detection and tracking in a real scene immediately.There are many difficulties in face detection because of human face's complicate structure, personal feature, expression changes, face rotation, shadow and so on. When refered to face detection in video sequences, real time detecton must been guaranteed and object occlusion must be handled. As a result, real-time tracking and update the motion model in time are the key problems to be solved.For a face tracking system in video sequences, the core includes two parts: human face detection and face tracking. Human face detection is the basic part in a face tracking system. Before tracking faces in video sequences, face must be detected first. Only locating the face position correctly, can we track the moving faces. With incorrectly extracted or half-baked faces, it may seriously affect the tracking operation, even can not track the object. As the development of fields in computer vision, object detection, more and more new methods are used for face detection and tracking. Nowadays, the AdaBoost algorithm based on cascade structure is considered as the the most efficive method for object detection. In this paper, AdaBoost algorithm is used for face detection in video sequences. After analysing the AdaBoost algorithm thoroughly, A modified AdaBoost algorithm is proposed to reduce the training time by at least 20 times, meanwhile, guaranteed the detection accuracy. To improve the detection efficiency based on cascade structure classifier, a revised detection algorithm based on skin color is proposed, which made real-time multi-view face detection possible in video sequences. The way using skin color was different from the previous methods, which is not based on morphology or connected domain. On the object tracking research, real-time tracking and detection rate must be guaranteed. CamShift algorithm is widely used for fast tracking in video sequences for faces are searchend in a local area instead of the whole image. Occusion is unavoidable in a scene where there are more than one object. Kalman filter is used for predicting the face lacation when occusion happens. The experiment result shows that the algorithm presented in this paper can handle multi-view face tracking in video sequences effectively.