Abstract

1. Introduction Crowd anomaly detection (CAD) is one of the most interesting topics among researchers in the fields of computer vision and intelligent video surveillance. With the increase in population density and number of people attending public events, detecting anomalies becomes very significant as this can aid in the prevention of any potential threats to public security. This paper provides a complete overview of crowd anomaly detection techniques which will cover the basics of crowd anomalies, different types, various deep learning frameworks, benchmarking datasets, comparative analysis of system performance, current challenges, and possible ways forward. The study focuses on three major learning approaches namely supervised, semi-supervised and unsupervised learning techniques. Additionally, the review highlights the advancements in current techniques used in crowd anomaly detection including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Generative Adversarial Network (GAN), Graph Convolutional Network (GCN) and discusses their performance using benchmark datasets including UCSD, UMN, ShanghaiTech and UCF-Crime datasets.Introduction

Surveillance systems through video cameras have now become an integral part of the safety infrastructure in modern cities. The ability to constantly analyze the crowd behavior — detecting panic, stampede, violence, or suspiciously loitering people — is essential in order to prevent disasters caused by crowds. Examples of such crowd-related disasters include the stampede that occurred on the New Year's Eve in 2014 in Shanghai, China, and the stampede that occurred during the 2015 Hajj pilgrimage in Saudi Arabia.

The conventional crowd monitoring process involves human analysts watching live video from several cameras at once. However, this method is ineffective due to its cost, non-scalability, and vulnerability to errors. Therefore, automated systems for detecting anomalies in crowded scenes utilize advanced technologies like machine learning and computer 

vision for constant video stream monitoring and immediate alerts for security agents.

With the development of surveillance technologies over the last decade, there have been remarkable advancements made in this area. This can be seen in the rising number of research articles on crowd

anomaly detection that increased from less than twenty published articles each year prior to 2015 to over 180 in 2022.

The organization of this paper is as follows: In Section 2, the concept of crowd anomaly detection is explained along with its various aspects. Section 3 provides a classification of various types of detection systems. Section 4 highlights advances made with deep learning techniques. Section 5 describes how various systems compare to each other when tested using benchmark databases. Section 6 provides publicly available datasets.