Improving Security Surveillance with Weapon Detection via Faster R-CNN

Abstract

Security remains a critical concern across multiple domains due to the increasing incidence of crime in densely populated areas, public events, sensitive zones, and remote locations. The detection and monitoring of abnormal activities have emerged as key applications of computer vision in addressing these challenges. Weapon or anomaly detection involves identifying rare, irregular, or unexpected patterns that deviate from established norms within visual data. With the growing demand for public safety and the protection of assets, surveillance systems have become integral to modern security infrastructures, enabling the analysis of suspicious behaviours and unusual events for effective decision-making. Object detection techniques rely on feature extraction and deep learning models to accurately recognize various objects, where minimizing false positives is essential to prevent inappropriate responses. Consequently, selecting a robust and efficient methodology is crucial for ensuring high performance. In this study, an automated firearm detection system is proposed using convolutional neural network-based models, specifically Single Shot Detector (SSD) and Faster R-CNN. The proposed approach is evaluated using two types of datasets, one comprising pre-annotated images and the other consisting of manually labelled data. Experimental results demonstrate that both models achieve promising accuracy, while their real-world applicability depends on an optimal trade-off between detection speed and precision.