Study On Face Tracking And Recognition

Face analysis is one of hot topics in the field of computer vision and pattern recognition because of the potential applications in many fields, such as identity authentication, surveillance, human-computer interface,multi-media and so on.. It includes face detection, tracking, recognition, facial expression analysis, modeling and animation. In this thesis, we focus on face tracking and recognition. The contributions of the thesis are:1. A robust 2-D head tracking is proposed which is based on histogram matching and shape constrain. First, an adaptive weighted histogram matching is used to estimate an initial position, in which an optimal method called mean shift is adopted to search matching path automatically. After histogram matching, a normalized gradient model of elliptical boundary is used to accurately track the head's position and scale size in a local range. Experiments demonstrate that it is a real-time and robust tracker.2. Since linear Principal Component Analysis (PCA) was successfully applied for face recognition, subspace analysis methods have been one kind of popular methods. In the thesis, a brief review of subspace analysis methods in face recognition is given.3. In order to get rid of the constrain of Probability Reasonable Model (PRM), kernel density estimation is presented to estimate the within-class conditional probability, and EM algorithm is adopted to estimate the radius of the kernel. Experimental results show that it can improve the performance of linear Principal Component Analysis (PCA) and Kernel based Principal Component Analysis (KPCA) in face recognition.4. Because it is inadequate for linear subspace analysis methods to describe the complex relations of real face images, such as pose, illuminant, expression variations. Kernel based Fisher Discriminant Analysis (KFDA) is proposed for face recognition, which combines the nonlinear kernel trick and Fisher Linear Discriminant Analysis(FLDA). Experimental results show that it can give higher accurate recognition rate than linear subspace analysis methods and KPCA.5. Based on the previous work, in order to further enhance the performance of KFDA in face recognition, the Cosine kernel function is proposed to replace the original polynomial kernel function, and feature vector selection is introduced to reduce the computational complexity, and the Nearest Feature Lines (NFL) classifier is combined. Experimental results show that the proposed method has an encouraging performance.