Abnormal High-Density Crowd Dataset

Anomaly detection within crowded environments is a key challenge in the computer vision and crowd behaviour understanding fields. Furthermore, anomaly detection within high-density crowds remains an insufficiently explored area. In this paper, we propose a novel abnormal high-density crowd dataset. The proposed dataset adheres to the same constraints as some of the benchmark datasets such as UCSD, UMN and Avenue dataset. These constraints include occurrences of both normal and abnormal behaviour and validated abnormal behaviour annotations. Additionally, this dataset includes footage of only high-density crowds, whereas benchmark datasets include low to medium density crowds. We have taken into consideration privacy issues, the veracity of annotations and pre-processing of the dataset. We evaluate the dataset against state-of-the-art crowd anomaly detection methods. The generated results indicate that training/testing these methods on high-density crowds decreases their detection performance.

I. Introduction

Crowd behaviour analysis is an important matter of concern, in terms of both safety and security. Crowd formations are present in streets, public events, concerts, airports, religious pilgrimages, marathons etc., these venues are in danger of crowd disaster occurrences. Video surveillance has been increasing in many environments to enhance security and prevent disastrous situations. Consequently, a substantial amount of data is generated from multiple sources, therefore overwhelming surveillance operators. Automation of crowd behaviour analysis with limited human supervision is required to enable smarter and safer environments. To achieve this, specific information is extracted from surveillance footage using computer vision tasks to automatically understand the behaviour of a crowd.