Uma abordagem dinâmica para detecção e seguimento de face em vídeos coloridos em ambientes não controlados

Abstract

Face detection in images is a growing branch of computer vision, regarding its potential for several applications. Identification, authentication and recognition of individuals are some tasks that are performed by systems that rely on such techniques. In this thesis, an approach is described to to detect and track a face in color uncontrolled videos. The Dynamic Support Vector Tracker (from the Brazilian Portuguese Seguidor Dinâmico com Vetores de Suporte - SDVS) framework here proposed combines face detection with target tracking, through integrating illumination compensation, skin color detection, Gabor features and a discrete Kalman filter, thus implementing an integrated system. Such architecture differs from others found in the literature for being able to detect and track faces in unconstrained outdoor videos, under real-world conditions, with different skin tones, tracking arbitrary poses of the face and for being capable of recovering failures in face detection. To validate the SDVS, tests were performed on videos from the Honda / UCSD and David Ross Video Databases, as well as videos captured at the Goiabeiras campus of the Federal University of Espírito Santo. These videos were categorized according to the degree of difficulty (challenge) to be treated, and the results of applying SDVS to them were compared with the correspondent results associated to a state of the art technique in order to evaluate the performance of the SDVS. The results suggest that the approach here proposed and the SDVS system implemented are validated.