Violence Content Detection Based on Audio using Extreme Learning Machine

Abstract

In this paper, we proposed an audio based violent scene detection system. As visual based approach has been widely used in identification of violent scenes from video data, audio-based approach; on the other hand, has not been explored as much as visual approach of the video data. In some applications such as video surveillance, visual scenes can be absent because of environmental situations. Also, in many approaches different systems are proposed for movies and real time videos. Therefore, we practiced the audio approach of video data to decide whether a video scene is violent or not from movies and real time videos. For this purpose, we propose an Extreme Learning Machine (ELM) method to detect video scenes as “violent” or “non-violent” using two types of datasets Standardized Media Eval VSD-2014 and other is Customized dataset for the same classifier. After successful training and testing, 85.7% accuracy is achieved by ELM for VSD-2014 dataset and 88.89% for Customized dataset respectively.