Study on Deep Learning Techniques for Finding Suspicious Violence Detection in a Video Surveillance

Abstract

Detecting suspicious visual objects is essential to applying automatic violence detection (AVD) in video surveillance. Continuous monitoring of objects or any unusual things is a tedious task. Learning about video surveillance is an emerging research problem in AVD applications. Deep learning is an intelligent and trustworthy technique for detecting or classifying suspicious data objects. It classifies suspicious video frames by modeling specific categories of videos. The current deep models convolutional neural network (CNN), convolutional long-term and short-term memory (ConvLSTM), AlexNet, VGG-16, MobileNet, and GoogleNet, are wildly succeeded in real-time violence detection with the input of video clips. This paper presents the findings of experimental studies for deep models using classification measures to demonstrate the models' efficacy for our AVD application. Benchmarked violence (V), non-violence (NV), and weapon violence (WV) video datasets are used in the experiment to describe the model's performance while classifying suspicious videos for public safety.