%0 Conference Proceedings
%4 sid.inpe.br/sibgrapi/2019/09.08.15.38
%2 sid.inpe.br/sibgrapi/2019/09.08.15.38.29
%@doi 10.1109/SIBGRAPI.2019.00010
%T Fight Detection in Video Sequences Based on Multi-Stream Convolutional Neural Networks
%D 2019
%A Carneiro, Sarah Almeida,
%A Silva, Gabriel Pellegrino da,
%A Guimarães, Silvio Jamil F.,
%A Pedrini, Helio,
%@affiliation Institute of Computing, University of Campinas
%@affiliation Institute of Computing, University of Campinas
%@affiliation Computer Science Department, Pontifical Catholic University of Minas Gerais
%@affiliation Institute of Computing, University of Campinas
%E Oliveira, Luciano Rebouças de,
%E Sarder, Pinaki,
%E Lage, Marcos,
%E Sadlo, Filip,
%B Conference on Graphics, Patterns and Images, 32 (SIBGRAPI)
%C Rio de Janeiro, RJ, Brazil
%8 28-31 Oct. 2019
%I IEEE Computer Society
%J Los Alamitos
%S Proceedings
%K Fight detection, convolutional neural networks, video analysis.
%X Surveillance has been gradually correlating itself to forensic computer technologies. The use of machine learning techniques made possible the better interpretation of human actions, as well as faster identification of anomalous event outbursts. There are many studies regarding this field of expertise. The best results reported in the literature are from works related to deep learning approaches. Therefore, this study aimed to use a deep learning model based on a multi-stream and high level hand-crafted descriptors to be able to address the issue of fight detection in videos. In this work, we focused on the use of a multi-stream of VGG-16 networks and the investigation of conceivable feature descriptors of a video's spatial, temporal, rhythmic and depth information. We validated our method in two commonly used datasets, aimed at fight detection, throughout the literature. Experimentation has demonstrated that the association of correlated information with a multi-stream strategy increased the classification of our deep learning approach, hence, the use of complementary features can yield interesting outputs that are superior than other previous studies.
%@language en
%3 paper.pdf