QuARCS: Quantum Anomaly Recognition and Caption Scoring Framework for Surveillance Videos

Abstract

Traditional surveillance video stream monitoring demands manual analysis, often leading to inaccuracies. While recent advancements have enabled automated analysis in surveillance video stream monitoring, challenges persist in achieving high accuracy and efficiency. Thus, an automated system is needed to monitor and report on video streams in real-time or retrospectively within surveillance networks, alleviating human error and inefficiency. Our paper, presents a comprehensive framework that integrates a hybrid quantum-classical anomaly detection system, a caption-generating model, and a novel Text-Driven Urgency Rating Model (T-DURM) trained using a newly created labelled dataset called UCFC-CUR which prioritises crimes based on their urgency. The hybrid classifier outperforms its direct classical counterpart by 7.7%. The aforementioned pipeline possesses the capability to identify anomalous occurrences from surveillance videos, generate a textual representation of the event, and assign a numerical value indicating the level of urgency associated with the specific anomaly. The hybrid anomaly detection model achieved an AUC of 82.80 surpassing the classical model’s AUC of 75.14. While the newly proposed T-DRUM achieves a R2 score of 0.982.