Reconhecimento de ações com histogramas de características visuais e contexto adicionado por tranferência de aprendizagem

Abstract

This thesis addresses the task of recognizing human actions in realistic videos based on their visual content. Such an ability has a wide variety of applications in specic settings, but this work is above all motivated by the idea that efective visual descriptors and models need to be provided in order to make current search engines better able tocope with the large amount of multimedia data being produced every day.An issue which has arisen from preliminary studies is the fact that to manually collect action samples from realistic videos is a time-consuming and error-prone task. This is a serious bottleneck to research related to video understanding, since the large intra-class variations of such videos demand training sets large enough to properlyencompass those variations. In this thesis, we propose an approach for this problem based on Transfer Learning (TL) theory, in which we relax the classical supposition that training and testing data must come from the same distribution. Our experiments with Caltech256 andHollywood2 databases indicated that by using transferred information from only four concepts taken from the auxiliary database we were able to obtain statistically signi cant improvements in classication of most actions in Hollywood2 database, thus providing strong evidence in favor of the presented solution. Such solution encompasses our main thesis, which can be summarized in two main contributions: a) it is feasibleto use TL techniques to detect concepts in realistic video action databases and, b) by using the transferred information, it is possible to enhance action recognition in thosescenarios.