Violence Detection in Videos Using Neural Networks.

Abstract

In this project, We explore and investigate strategies in this project in order to detect violence to totally dismantle the current situation and foresee upcoming trends in inves tigation of violence. We would offer a complete evaluation of video violence detection challenges outlined in cutting-edge research in our systematic research. This work will explore and identify any outstanding concerns in the current situation or subject, as well as technologically advanced methodologies in video violence detection, data sets for constructing and training video in real time violence frameworks for detection, and debates and identification of outstanding issues. We analyzed 80 research papers chosen among 154 after the Phases of identification, screening, and eligibility in this study. We starts with rapidly explaining the fundamental concept and challenges of video-based violence detection; next, we divide current solutions based on their methodologies, they are divided into three categories. There are traditional methodologies, end-to-end deep learning-based methods, and machine learning-based ways. Finally, we offer and assess video-based violence identification algorithms for testing their efficacy. Furthermore, we describe the open difficulties in video violence identification and assess its future trends. Hockey Fight videos data set has binary classes such as violence and no violence with 1000 videos data, where 800 video’s for training and 200 video’s for testing the model performance. The process of extracting features, classification, and pre-processing pro cess are the processes that were used in this study. Classification with a success of 93.7 percent accuracy was produced by VGG16-LSTM, which outperformed 93.4 per cent from VGG16-LSTM, 92.3 percent from EfficientNetB0-GRU, 92.1 percent from VGG19-GRU, 90.9 percent from VGG19-LSTM, 90.4 percent from InceptionV3-LSTM, 90 percent from EfficientNetB0-LSTM, 89.1 percent from InceptionV3-GRU, 83.9 per cent from Resnet50-LSTM and 82.6 percent from Resnet50-GRU for the classification of Hockey fight videos. According to other evaluation performances such as sensitivity,