Automatic violence detection in video surveillance is crucial for social and personal security. Due to the massive video data produced by surveillance cameras installed in different environments like airports, trains, stadiums, schools, etc., traditional video monitoring by humans operators becomes inefficient. In this context, develop systems capable of detect automatically violent actions is a challenging task. This study describes a method to detect and localize violent acts in video surveillance using dynamic images, CNN's, and weakly supervised localization methods. Experimental results demonstrate the effectiveness of our approach when applied to three public benchmark datasets: Hockey Fight [1], Violent Flows [2], and UCFCrime2Loca1 [3]. © 2020 IEEE.

This work was supported by grant 234-2015-FONDECYT (Master Program) from Cienciactiva of the National Council for Science,Technology and Technological Innovation (CONCYTEC-PERU).