Research Of Face Detection And Tracking Algorithms Based On Boosting

Face detection and tracking has great application prospects in the field of secure authentication, human-computer interaction, the pursuit of criminals, and video conferencing. Therefore, face detection and tracking have become a hot research topic of the current pattern recognition and artificial intelligence. The purpose of face detection is to judge whether there is a face in an image or image sequence, if exists, it returns the position and space distribution. Face tracking refers to determine the trajectory of the face and size changes in the input image sequence for further analysis of the movement parameters of the face. This dissertation draws on theories and techniques of state-of-the-art at home and abroad, combined with their own innovation, carried out a series of algorithms and experiments.The main work in the dissertation is listed as follows:(1) Reviewed the development history and study actuality of the face detection and tracking technique and illustrated significance of the dissertation.(2) In face detection, face detection algorithm based on skin color model and face detection algorithm based on AdaBoost are achieved. Face detection algorithm based on AdaBoost, which is of high acuracy, but sometimes it brings false detection. Face detection algorithm based on skin color model is easy, real-time and can effectively eliminate non-color target. Therefore, the dissertation puts forward the combination of Adaboost and the skin color model. The test results verified by the skin color detection method, which can decrease the false detection rate from4.2%to3.8%.(3) The online boosting algorithm is realized in the dissertation. In the tracking experiments, the first target to be detected defined as the positive samples to the learning classifier, followed by cycle to complete the face detection process in order to achieve tracking. However, the method is prone to drift tracking.In order to improve the diversity of the classifiers, the dissertation consider the distinction between classifiers, which makes the tracking more robust. Tracking accuracy increased from90.3%to94.6%, at the same time, miss rate decrease from6.9%to3.5%. (4) A face tracking algorithm which embedded the online boosting into particle filter framework is proposed. First the particle filter algorithms based on region-match tracking are studied, which uses a template matching method to calculate the similarity.On the basis of the algorithm, this dissertation uses the eigenvalue that detected by online boosting strong classifier as a template, compared to calculate the similarity with the corresponding eigenvalues of each particle, in order to calculate each particle weight. The experimental result show that the drift phenomenon of tracking is improved and tracking is more accurate.The algorithm automatically detects the face, and update the template every5seconds automatically, which achieved automatically face detection and tracking.