Face Position Detection System For Stereoscopic Display

With the development of the face detection technology, different kinds of the human-machine interaction systems have been applied in our life. people put forward higher requirements for the human-machine interaction technology to realize the facial features localization based on the face position detection. As an important field in machine vision, face position detection is involved in several fields, such as digital image processing, machine learning, pattern recognition and so on. The face position detection is the key part of the human-machine interaction technology which needs the features of real-time, high precision and so on.For the current human-computer interaction system, according to whether the system contacts with the observers, the systems can be divided into intrusive human-machine interaction system and non-intrusive human-machine interaction system. Due to non-intrusive human-machine interaction system does not contact the observers, so the experience of the observers is more comfortable and natural. Therefore, non-intrusive human-machine interaction system has more a good application prospect. As one kind of non-intrusive human-machine interaction system, non-aided stereoscopic display system uses the built-in camera device, and the pixels which the facial features contain in the video is limited. In order to accurately, stably position the facial features of the viewer in real-time, this paper presents a system used for facial features localization combined with face detection, facial features localization and target tracking at present, using the order of skin detection, face detection and facial feature localization to complete the cascade system with many kinds of algorithms.In order to improve the expansibility of the system and reduce the maintenance cost of the system, a system of the auto stereoscopic display is designed, which is divided into three layers:hardware layer, middle layer and view layer. By isolating different modules, the modules are decoupled and the system is easy to maintain.In order to ensure the accuracy face position detection and facial features localization and improve the running speed of the whole system, the skin color detection is the first part of the algorithms. By using the elliptical boundary model, cluster the pixels of skin color, and separate the candidate regions from the background to reduce the number of detection windows. And through according to the different light intensity to adjust the parameters, improve the robust of the algorithm and the efficiency of classification.In facial feature localization, ASM(active shape model)algorithm is adopted. Using 68 feature points of the model, the face global shape model and the local texture feature are established, and the distribution of the feature points and the texture information around the feature points are described. The distance between the targets is calculated by using the Markov distance when determining the position of the feature points, and the importance of the dimension of the smaller variance is highlighted. After several rounds of iteration the optimal location of the feature points are found and precise position of the facial feature is achieved.In target tracking, TLD(Tracking-Learning-Detection)algorithm is adopted. Compared with the traditional tracking algorithm, by adding a learning module based on detection module and tracking module.The detection module cascade the patch variance, ensemble classifier and nearest neighbor classifier to detect the target. The tracking module uses the Lucas-Kanade optical flow algorithm to track the target. The positive and negative samples are obtained when the target is detected and tracked successfully. The learning module uses the semi supervised learning to learn the positive and negative samples and adjusts the parameters of the detection module and tracking module so that a long time target tracking is achieved.The algorithm proposed in this paper achieves the real-time and high precision requirements of the non - aided stereoscopic display system, and algorithm is robust for the face rotation, partial facial occlusion of the observers and different complex circumstances such as the change of external illumination condition and so on.