Anomaly detection on the edge using smart cameras under low light conditions

Abu Awwad, Yaser 2023. Anomaly detection on the edge using smart cameras under low light conditions. MPhil Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Yaser Abu Awwad, MPhil, Thesis) - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (31MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (298kB)

Abstract

The vast number of cameras utilised in smart city domains is becoming increasingly prominent and notable for monitoring indoor and outdoor areas such as buildings and road traffic as well as in rural areas (e.g. farms) in order to deter thefts of farming machinery and livestock, besides monitoring workers to guarantee their safety. In addition, to detect anomalies meant as Identifying an unusual occurrence that does not adhere to the nature and regulations. However, detecting anomalies becomes much more challenging in environments with low lighting and poor visibility conditions, such as at night (when a scene is entirely dark) and partially dark during dusk and dawn, resulting in obtaining inefficient outcomes to recognise events leading to an increase in false positives detection. Thus, this research aimed to identify objects (referred to as Anomalies) in low-light settings with the assistance of pre-existing methodologies in image enhancement and object recognition on resource-constrained devices (referred to as Nodes). Rather than focusing on enhancement methods for image quality comparison purposes or developing a novel approach, the main goal is to exploit existing methods to boost the detection stage’s accuracy. Further, a lightweight classification algorithm is proposed to differentiate (1) Bright scenes captured in the daytime from low-light ones and (2) To distinguish between low-light scenes that might differ in their darkness levels and incorporate additional factors such as noise. Therefore, images with insufficient light are enhanced by multi-enhancement networks, where the optimal one is chosen based on the input image features and characteristics. The results demonstrated an increase of 25% & 3% in the detection accuracy on the ExDark database. Moreover, the classifier could discern between bright and dark scenarios achieving an accuracy of 85.24%. Finally, the proposed classification, enhancement, and detection stages were implemented on the resource-constrained devices, demonstrating efficiency and resilience, retaining high performance and time-response roughly (1 second) across all phases.

Item Type:	Thesis (MPhil)
Date Type:	Completion
Status:	Unpublished
Schools:	Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Date of First Compliant Deposit:	7 December 2023
Last Modified:	08 Dec 2023 10:48
URI:	https://orca.cardiff.ac.uk/id/eprint/164603

Actions (repository staff only)

Edit Item