FusedVision: Advancing Anomaly Detection in Surveillance Videos with Branched Networks

Document Type



Surveillance videos play an important role in ensuring the safety and security of public and private spaces, buildings, and critical infrastructure. However, manually monitoring these videos is a challenging and time-consuming task, and there is a need for automated systems to detect anomalies and potential security threats. Anomaly detection is a critical task in surveillance video analysis, as it helps in identifying abnormal behavior and events in real-time. In this work, we propose a novel approach to detect anomalies in surveillance videos using computer vision techniques. In recent years, object-centric anomaly detection methods have gained significant attention due to their higher performance and object- targeted approach. These approaches typically involve using an object detector to crop all objects appearing in the frames and then training a learning model using the objects and their relations. In the present study, a unique approach for object-centric anomaly detection is proposed, utilizing a branched network architecture comprising an object de- tector and a normalcy learning model to segregate anomalies from the normal data. In order to attain peak performance, an in-depth investigation was conducted to analyze the most effective fusion mechanism for the two branches. The proposed framework demon- strates superior performance by leveraging the strengths of both models, which results in reduced false alarms and improved localization/detection of anomalies. The general applicability of the proposed framework is illustrated by integrating it with five existing state-of-the-art anomaly detection methods and assessing its performance on three dif- ferent datasets: ShanghaiTech, Avenue, and Ped2. The experimental results show that our proposed approach noticeably outperforms existing methods, demonstrating its effec- tiveness in detecting anomalies across a range of contexts. The efficacy of the proposed work implies its potential applicability across diverse domains, including but not limited to surveillance, security, and anomaly detection.

First Page


Last Page


Publication Date



Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfillment of the requirements for the M.Sc degree in Computer Vision

Advisors: Dr. Abdulmotaleb Elsaddik, Dr. Karthik Nandakumar

Online access for MBZUAI patrons