Abel Diaz Berenguer
Paired supervised learning and unsupervised pretraining of CNN-architecture for violence detection in videos
Recognizing violence in crowded scenes is a major challenge for automatic video surveillance. We propose a shallow CNN that is pretrained using an unsupervised strategy. Afterwards, the pretrained parameters are fine-tuned to extract intermediate frame representations, which are subsequently aggregated to obtain meaningful video representations to recognize violence in footage. We validated that automatic crowded scenes analysis can benefits from domain specific representations while shallow CNN can achieve competivtive performance.
Received the B.Sc.Eng. degree in informatics sciences and M.Sc.Eng. degree in applied informatics from the University of Informatics Sciences (UCI), Havana, Cuba in 2009 and 2014 respectively. He is currently working towards a PhD degree at the Audio-Visual Signal Processing Laboratory of the VUB Electronics and Informatics Department. His current work focuses on automatic human behavior analysis and smart video surveillance.