Video Surveillance Anomaly Detection: A Review on Deep Learning Benchmarks

Many surveillance cameras are mounted in sparse and crowded indoor and outdoor areas to monitor and detect various patterns of human behaviors and anomalies in the public and private sectors. The continuous streams from these cameras produce an enormous amount of graphical data. This data featuring anthropometric, multi-view and large-scale variations, perspective distortions, serious occlusions, cluttered environment, dynamic behaviors, diverse anomalies which requires a high computational cost and particular hardware resources. These factors, combined with the sheer volume of data, make manual analysis impractical, especially for real-time monitoring. To address these challenges, researchers have been developing automated visual surveillance systems. These systems have evolved from classical approaches to cutting-edge Deep Learning (DL) methods, i.e. CNN (Convolution Neural Networks), transfer and sequential learning techniques, transformer models, YOLO (You Only Look Once) algorithms and auto-encoders. The significant features of DL methods, such as weight sharing, parameters reduction, large-scale network implementation, automated feature extraction, ability to handle time-series data, real-time performance capabilities and computational efficiency are the reasons for promising results. Despite significant progress, there is still no definitive solution for Video Surveillance Anomaly Detection (VSAD) that can handle real-time, large-scale datasets with diverse anomalies and behaviors. This study aims to provide a comprehensive analysis of SOTA (state-of-the-art) DL approaches for VSAD. It will examine the limitations and trade-offs of various learning models and datasets used in these approaches, contributing to the ongoing development of more effective automated surveillance systems.