Real-time crime monitoring system using deep learning for weapon, behavior, and facial detection

This study presents an efficient crime monitoring system (CMS) that leverages the UCF crime dataset, which contains actual surveillance footage of various criminal activities. The CMS integrates real-time video surveillance with deep learning (DL) models for crime detection, including weapon identification, misbehavior detection, and facial recognition. It utilizes hyper capsule networks (H-CapsNets) for identifying visible and concealed firearms, 3D convolutional neural network (C3D) for detecting anomalous behavior, and deep convolutional neural network (CNNs) for facial feature extraction and matching with a suspect database. Real-time detection is achieved with low response times—generating alerts in <1.5 s for weapon detection and 2.2 s for misbehavior detection, ensuring rapid intervention. Data augmentation techniques, including rotation, scaling, flipping, and panning, enhance model generalization, resulting in 93.5% accuracy, 92.8% precision, 91.7% recall, and 92.2% F1-score. The CMS demonstrates excellent real-time performance with low false positive (FP, 3.1%) and false negative (FN, 2.7%) rates, showcasing its potential for practical deployment in public safety systems.