Investigating human detection and tracking techniques in unconstrained surveillance videos

Abstract

The goal of computer vision is to let computers understand a scene through the eyes of a camera. In recent years, the number of surveillance cameras installed to monitor private and public places has increased dramatically. Human detection and tracking in surveillance videos play a critical role in ensuring public safety, security, and efficient management of crowded places. Visual surveillance systems rely on advanced computer vision techniques to detect, identify and follow individuals in video frames. While significant progress has been made in this field, existing human detection and tracking methods often struggle to perform reliably in unconstrained environments that are characterized by unpredictable lighting, dynamic backgrounds and occlusions. These challenges limit the effectiveness of current approaches in real-world applications. This thesis investigates state of the art human detection and tracking methods by focusing on their performance in unconstrained surveillance videos from Sydney train stations and the well-known MOT20 benchmark. The research systematically evaluates these methods under various challenging scenarios and provides insight into their abilities, limitations and opportunities for improvement. The study also investigates the application of these methods in public safety.

Date of Award	2025
Original language	English
Supervisor	Roland GOECKE (Supervisor), Dat TRAN (Supervisor) & Ibrahim RADWAN (Supervisor)