Multi-View Face Detection And Tracking

Face detection and tracking is becoming a significant research aspect in computer vision and pattern recognition. With the improvement in intelligent video surveillance, human-computer interaction and video conference fields, automatic face recognition in complicated scenes is wildly required. Thus, multi-view face detection and tracking attract a widespread attention for multiple pose face detection. Meanwhile, face as a typical non-rigid object, research in multi-view face detection and tracking field can greatly improve object detection and tracking technology.After surveying on recent research achievement, this thesis mainly study on key technologies of multi-view face detection and tracking, including multi-view face detection algorithm, tracking method based on online learning, multiple objects tracking system and accurate eye location algorithm. The main contributions are listed as follows.(1) Based on the prior knowledge of multi-view face detection, i.e. the structure of multi-view classifier, AdaBoost algorithm, Vector Boosting, and weak classifier function, a novel incremental learning multi-view face detection method is proposed based on Width-First-Search tree. Compared with state-of-the-art algorithms, the proposed method shows a better performance on detection rate, false alarm and generalization.(2) A novel online learning algorithm (Semi-MIBoost) is proposed, which combines the robustness of online semi-supervised learning and the flexibility of multiple instance learning method. Thus, the algorithm can effectively tackle the tracking drift problem.(3) Based on the proposed online Semi-MIBoost algorithm, a new multiple objects tracking framework is proposed. The adopted sample selecting strategies allow the tracking system efficient using the offline multi-view face detection model and rapid self-updating. Experimental results show the better performance of the proposed method comparing with state-of-the-art methods on several benchmark videos, which contain several object rapid scale change and full occlusion scenes.(4) A novel coarse-to-fine eye location method is realized, which includes Integral Projection Function, rough single eye location algorithm based on AdaBoost, accurate eye location method based on isophotes center map, and eye pair validation algorithm. The experimental results demonstrate that the proposed method has better location accuracy and less computation, especially under illumination changes circumstance.