Research And Application On The Face Detection And Tracking Technology

Automatic face detection and tracking is an important and front-line subject in the field of man-machine interaction. Recently, this subject has become an active research topic for many organizations all over the world due to promotion from intelligent surveillance, videoconference and video-retrieval. In this paper, we have made a deep study on the automatic face detection and tracking in color image sequences. The major works of this thesis including as follows:1. After researching the appearance and development of face detection, we study the Adaboost algorithm particularly and introduce the concept of 'Haar rectangle features', 'Integral Image' and 'Cascade'. Inspired by the 'Integral Image', we propose a conception of 'color probability Integral Image' and a new method of face detection on the basis of Adaboost algorithm and skin pre-detection through sub-windows. We integrate the skin model into the cascade detect process of the Adaboost algorithm ingeniously. Experiments demonstrate this algorithm can reduce the false-alarm probability and elevate the accuracy.2.We propose a new face tracking algorithm on the basis of improved CAMSHIFT algorithm and face detection. At first we detect the face region quickly using the Adaboost algorithm. Then a color histogram model is constructed after the more precise area of the face is get due to the Canny edge detection. Besides the CAMSHIFT algorithm, Kalman filter and human's face feature is also used in our system. The experiment results demonstrate that the system is not only real time but also handle well the complex situation such as occlusion, disturbing by skin colored object and fast moving.3. We propose and realize a face tracking method on the basis of particle filter, skin features and ellipse restriction after studying the common theory of particle filter in order to handle the tracking under complex background. Experiments demonstrate good results can also be gained under complex background using our algorithm.