Real-time Anomaly Detection and Multi-class Classification in Surveillance Videos Using Deep Learning

Abstract

With the escalating crime rates in the school and colleges, surveillance cameras have become instrumental in monitoring and detecting anomalies happened in the campus to reduce crime rates. This study proposes a novel deep learning-based approach for real-time anomaly detection and multi-class classification in surveillance videos. We employ a Convolutional 3D Neural Network (C3D) model for extracting features and detecting probabilistic anomalies within video frames. Subsequently, a Simple Recurrent Neural Network (RNN) is trained to classify detected anomalies into one of thirteen distinct categories, including Abuse, Arson, Assault, among others. Our approach outperforms previous methods, achieving an accuracy of 83.67% for anomaly classification on the UCF Crime dataset. This work demonstrates the feasibility of automatically detecting and classifying anomalies in real-world surveillance scenarios, potentially enhancing public safety and well as school/college campus safety and facilitating prompt response measures.