Skip navigation

Dynamic image crowd representations for improved anomaly detection using generative adversarial networks

Dynamic image crowd representations for improved anomaly detection using generative adversarial networks

Mahmoud, Samar (2021) Dynamic image crowd representations for improved anomaly detection using generative adversarial networks. PhD thesis, University of Greenwich.

[img]
Preview
PDF
Samar Mahmoud 2021 - secured.pdf - Published Version
Available under License Creative Commons Attribution.

Download (49MB) | Preview

Abstract

Crowd formations are inevitable in many environments, and hence planning for, and managing crowds are integral parts of city and event planning. Effective analysis of crowd behaviour and anomaly detection has the potential for more efficient management and is a building block for smart environments. Closed-Circuit Televisions (CCTVs) capture vast footage and are an important information source, some of which contain images of crowds of high density. However, relying on the typical manual surveillance systems for detecting anomalies (any behaviour outlying from established normalcy) in crowds presents complications concerning accuracy and computation power. This research intends to advance the automation of anomaly detection within medium and high-density crowds. Using crowd behaviour analysis methods, anomaly detection is applied to recognise occurrences of anomalous behaviour within crowds. An anomaly within the behaviour of the crowd is detected by analysing crowd footage with the use of deep vision algorithms. Results obtained from the processing of video data can be used to understand the overall scene and discriminate between normal and abnormal behaviour within a crowd.

Application of crowd anomaly detection has improved recently, however, the algorithms currently being used are usually time-consuming, computationally heavy, or require high power consumption. Amongst the work reviewed, both handcrafted approaches, as well as a variety of neural network approaches suffer from a lack of a definition of what “abnormal” behaviour is. Benchmark datasets used to train/test these methods lack sufficiently rich enough data to define anomalous behaviour. Therefore, abnormal events are considered as any events that deviate from the defined normal. Furthermore, state-of-the-art methods also present limitations of applicability to high-density crowds. High-density crowds are not targeted as much due to their difficulty in application. A key contribution of this research addresses this issue with the creation of a public anomalous high-density crowd dataset. The high-density dataset named Abnormal High-Density Crowd (AHDCrowd) has been utilised in training and testing the state-of-the-art crowd anomaly detection methods to evaluate their anomaly detection performance on high-density crowds.

Another key contribution of this research is a novel approach to crowd behaviour anomaly detection. Various dynamic image representations are used as an alternative to optical flow extractions for temporal development features extraction. The features are used in conjunction with image-to-image translation using CGANs (Conditional generative adversarial nets) for anomaly detection within crowds, and the proposed framework is evaluated on benchmark datasets as well as the AHDCrowd dataset. The applied experiments evaluate the effectiveness of utilising various types of dynamic image representation for crowd anomaly detection. The experimental results obtained have demonstrated the efficacy of this approach compared to the
state-of-the-art crowd anomaly detection methods.

Item Type: Thesis (PhD)
Uncontrolled Keywords: Surveillance, crowd behaviour, crowd anomaly detection,
Subjects: Q Science > QA Mathematics
Faculty / School / Research Centre / Research Group: Faculty of Engineering & Science
Faculty of Engineering & Science > School of Computing & Mathematical Sciences (CMS)
Last Modified: 14 Jul 2022 15:57
URI: http://gala.gre.ac.uk/id/eprint/37028

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year

View more statistics