AI-Powered Real-Time Surveillance: An Intelligent Threat Detection System for Security Enhancement

Abstract

Intelligent real-time surveillance systems are upending the status quo in high-risk and complicated contexts due to the growing need for advanced solutions to improve public safety. But their performance is frequently hindered by insufficient training data diversity, class imbalance, and limited generalization due to the varying environmental conditions. The YOLOv8 object detection model is the foundation of this study's AI-powered surveillance framework, which integrates hybrid dataset construction with a data preparation workflow that includes image preprocessing, data augmentation, dataset splitting, and SMOTE-based dataset balancing applied only to the training subset prior to YOLOv8 training. To increase environmental variety and model generalization, a hybrid dataset was created by merging the Microsoft COCO dataset and the CCTV Surveillance Image Dataset. Additionally, in order to enhance the representation of minority object categories prior to YOLOv8 training, SMOTE-based dataset balancing was only a