A Two-Stage Spatiotemporal CNN-YOLOv9 Framework for High-Precision Real-Time Abandoned Object Detection in Public Surveillance Videos

Maintaining security is of prime importance in public spaces such as markets, train stations, and airports. Such situations demand reliable and advanced automated surveillance systems. This research paper presents an intelligent video surveillance system for abandoned object detection in public environments. The overall architecture of the proposed systems consists of two stages. In the first stage, Convolutional Neural Networks (CNN) is used for real-time suspicious object recognition, efficiently classifying detected objects into suspicious and non-suspicious categories. In the second stage, the You Look Only Once (YOLOv9) algorithm, enhanced with a Kalman Filter for object tracking, performs precise spatial detection, and temporal persistence analysis is employed for abandoned object detection. Extensive experimental results on the PETS2006 and ABODA datasets—split into 70% training, 15% validation, and 15% testing—demonstrate outstanding performance, achieving up to 99.81% accuracy, 99.67% precision, and 99.01% recall, on benchmark datasets, with potential variations in unconstrained real-world deployments. The proposed approach significantly advances video-based threat detection and automated security monitoring systems, offering high detection accuracy, robust real-time performance, and strong adaptability to complex surveillance environments.