Abstract:
Detecting Violence in videos automatically and fast is very much essential as it's neither possible nor
feasible to continuously monitor a huge database of videos. It's critical to quickly determine violence as it
can be crucial in saving lives in real time violence cases in public places or fast detection among millions
of files on the internet to flag content, etc. Deep learning techniques have been proven to do well at various
tasks as compared to traditional algorithms but require many more resources like huge memory, high
computational power, etc. There is a great need to design deep learning architectures which are more
suitable to work in a constrained environment. We propose an architecture having Convolutional Neural
Network (CNN) for extracting the spatial information, which works as a feature extractor followed by a
Recurrent Neural Network (RNN) specifically Gated Recurrent Units (GRUs) to learn temporal cues. Our
contribution is proposing an architecture which is simple yet powerful. It uses very fewer parameters
(0.85million) without any degradation in performance as compared to state of the art results on benchmark
datasets like Hockey Fight, Violent Flows, and Movies. The simplicity of the architecture makes it suitable
for low constraint environments having the low computational power and less memory like mobile
devices, smart watches, etc.
