A combined skin model and feature approach for tracking of human faces

In this paper, we propose a face detection framework that combines both feature, and skin pixel approaches, while making the framework self adaptive which is important for non controlled environmental conditions. The framework uses skin color information to reduce the search space for faces by localizing the probable skin regions using a mixture of multivariate Gaussians whose parameters are first estimated using the Estimation Maximization (EM) algorithm. Then, feature based classification differentiates face related pixels from other skin regions and objects with close intensity values. A novel approach for classifying faces using a structure of cooperating neural networks, for which learning parameters are generated using Adaboost learning method is proposed. In addition, a new approach is also proposed for training the neural network with reduced space Haar-like features instead of working with image pixels themselves. Principle component analysis was used to find the aspects of features that are crucial for detection. The features dimensionality was reduced by nearly 90 percent, hence improving radically the training time. When adequate parameters are chosen, the system yields face detection characteristics that outperform the best existing algorithms (such as the one proposed by Viola and Jones) in terms of accuracy. Finally, the parameters of the mixture of Gaussians model are updated based on the results of the classification testing results to increase its robustness against illuminations and other external environmental changes, as well as reducing even more the search space.