Face detection and tracking base on video stream is an important research topic of computer vision and pattern recognition, also, it is a key technology in the field of facial information processing, which has become the facus of researchers on related subject. Automatic face detection and tracking technology is of great commercial value and application prospect in the fields of intelligent safety surveillance, human-computer interaction of video conference, automatic face recognition, identification, virtual reality. With great efforts taken by researchers all over the world, face detection and tracking technology has been developed greatly and acquired for practical effection under controlled conditions, however, due to the uncertain facial model, various noise interference during the digital image acquisition processing, practical environmental restrictions and theoretical research of immature, there are still enormous chanllenges, which bring great challenges to the performance of facial detection and tracking algorithm. Aiming at these confines of application condition, main research works are focused on face detection and tracking technology, comprehensive analysis of domestic and international's development in the fileds of face detetion and tracking, detailed study and in-depth analysis of the main algorithms. In order to satisfy the requirement of practical applications, our objective is to build a complete face detection and tracking system. The main research innovations and contributions are summarized as follows:1.We give a brief review for the basic knowledge of AdaBoost algorithm, then describe the process of how to construct weak classifier, strong classifer and cascade classifier in details based on the Harr–like features. For the process of studying and implementing the training and testing,finally, we improve the traditional AdaBoost algorithm as follows: (1) Optimizing the multi-scale detection algorithm to merge the overlapped face region in th images ; (2) Select different sample sets, to train frontal face classifier and profile face classifier respectively, then cascade them to get a new classifer with multi-level structure to solve the problems of multi-pose face detection. Large number of experimental data and classification results show that our system is able to detect most profile faces and the detection rate has been significantly improved. 2.By using radical template-based techniques to estimate the radial rotation angle, we effectively solve the problem in face detection of in-plane rotation. By cascading multiple classifiers, we successfully solve the problem in face detection of out-plane rotation.3.We deeply study the Mean Shift theory and its application in the field of target tracking, then anlysis the process of CamShift algorithm, which use Mean Shift as hardcore. Finally, we improve the CamShift algorithm as follows:(1) Using AdaBoost face detection algorithm to achieve automatic initialization of faces during the tracking process by CamShift algorithm; (2) In order to give attention to speed and effective of face tracking process, we combine the process of face detection with face tracking.Use the results of detection as the basis input of tracking while detecting the result of tracking to make sure whether the result of tracking is correct or not.(3) Considered classic CamShift algorithm used for face tracking when there being large color regions in the scene easily leading to inaccurate tracking, an improved CamShift algorithm integrating the template matching algorithm was proposed. We make the success of template matching in color probability distribution as iteration termination condition of CamShift algorithm can well figure out the trouble caused by color regions in face tracking. In order to reduce tracking errors and ensure maximum real-time algorithm, during the template matching process we scale the template and angle to rotate, and make the matching process only in the upper part of the search window.4.According to the above the theory and analysis, we develop and implement a automatic face detection and tracking system, using Visual C plus plus 6.0 software platform. Plenty of experiment results show that algorithm in this thesis obtains nearly ideal result in the field of detection rate false alarm rate, detection speed, and is a complete, robust, efficient face detection tracking algorithm with great all around performance. |