Sistema de reconhecimento de faces em vídeo.

Abstract

Automated face recognition is a technological area whose importance grows with the creation of security and monitoring systems in response to terrorist attacks. Its utility also lies in the areas of entertainment and commerce, where it allows for the automatization and simplification of processes. This work approaches the application of Support Vector Machines (SVM) to the task of automated facial recognition in video. Such process can be divided in three main parts, detection, feature extraction and recognition. Those three are further detailed in this work, although the main focus is in the problem of recognition and the application of SVM to it. The function of the SVM is to classify correctly a detected face from face images obtained from video. For that end, three multi-class classification methods based in SVM are studied, respectively: one-against-one, one-against-all and DAGSVM (Directed Acyclic Graph SVM). For face detection the Face Detector Viola-Jones method is explored, whose implementation called Haar Feature-based Cascade Classifier was used in the development of the recognition system. Specifically was utilized the version implemented in the Open Source Computer Vision (OpenCV) library. For the feature extraction step, preprocessing methods were used against the whole face image. For the recognition stage the LIBSVM library was used. It contains the implementation of the training and prediction methods for the SVM binary classifier. Tests using The AT&T Database of Faces, showed that DAGSVM was superior to the alternatives.