Assessing the Performance of Video Classification Models in Identifying Violent Actions

Abstract

Abstract: Manual video surveillance takes time and is prone to human error. Thanks to machine learning and deep learning, video analytics automates these processes. Video classification is an important aspect of video analytics. It is critical to detect violent behavior in recordings for surveillance, crime prevention, and public safety. Video classification algorithms are capable of detecting violent acts in video data. This paper introduces readers to the many approaches to video classification and describes the underlying network structure of the 3DCNN, ConvLSTM, and LRCN models, which are commonly used for video classification. Additionally, the models' implementation results were compared in order to conduct a comparative performance analysis of the models for the task of violent action classification. When it comes to classification, F1-Score, AUC score, and accuracy are useful metrics for evaluating models, and they were compared. We discuss some of the difficulties in violent action classification, as well as some potential future opportunities and new perspectives on how to address them and improve the system.