VioLess: A Vision based Violence Detection System

Abstract

For the past decade terrorism has been on the rise, growing at an exponential rate. There have been countless measures that are put in place to stop or at least minimize its effects. With the digital world proposing artificial intelligence based solutions to almost all major world problems, we aim to provide a system with a similar intent. The main objective of this system is to detect and localize unusual objects, suspicious behaviors or irregular events in a scene. The proposed method detects anomalous regions in a video. It codes video as a compact set of spatio-temporal video volumes while considering uncertainties in the grouping process and their spatio-temporal arrangements using a probabilistic framework to calculate the likelihood of the regions in the video. Our major target events would comprise of crowd analysis, video rating, forensics and violent events among people mostly, including kicking, punching, displaying of weapon, etc. The approach that we followed was to detect violence by using machine learning algorithms and by using deep neural networks. The major challenge we faced was the lack of resources for performing such high computationally demanding tasks and the lack of a local dataset. There are no such products in market at the moment which detect violence, we’re providing the world’s first deep learning based solution for violence detection. We ended up achieving a tool which can detect violence with more than 95 percent accuracy. We have currently focused on the violence part only, for the future we’re hoping to make this a full video analytics tool with person identification, number plate recognition etc. A product which eliminates the need of any human resource at all for all the surveillance purposes. Violent scenes provide as good a field of work as any, and with security threats becoming more of a worldwide concern, we aim to provide an effective solution despite of the numerous challenges faced in the area. Possible challenges includes the vast variations in video features that have to be considered e.g. video formats, camera angles etc., with inclusion for possible testing on all major video types. Moreover, in case of violent scenes, no major research has been recently done on actual real life scenes, primary focus has been on movies and fictional environments. This project requires us to set a concise differentiating line between violent and non-violent scenes.