FIREARMS MONITORING TECHNIQUE

Abstract

Violence in any form is a serious concern in modern society and continues to threaten the safety of innocent people. One common form of violence involves the use of firearms, which has become a global issue and poses major challenges for law enforcement agencies. Many firearm-related incidents occur in urban and semi-urban areas. Today, governments and private organizations widely use CCTV-based surveillance systems for monitoring and security purposes. However, manual monitoring of surveillance footage requires significant human effort and is often prone to errors. In contrast, automated smart surveillance systems offer better scalability and reliability for detecting violent activities. This paper focuses on demonstrating how deep learning techniques can be applied to detect firearms, particularly guns, in surveillance footage. The study uses advanced detection models such as Faster Region-Based Convolutional Neural Networks (Faster RCNN) and Efficient Det-based architectures to identify guns and human faces. An ensemble detection approach is used to enhance the performance of the models by combining multiple detection outputs at the post-processing stage. Techniques such as Non-Maximum Suppression, NonMaximum Weighted, and Weighted Box Fusion are applied to improve detection accuracy. The paper also presents a comparative analysis of different detection models and their ensemble combinations. These methods can help law enforcement agencies quickly gather information about possible threats and take preventive action. Additionally, the same approach can be applied to analyze social media videos for firearm-related content. The Weighted Box Fusionbased ensemble detection method achieved mean average precision values of 77.02%, 16.40%, and 29.73% for mAP0.5, mAP0.75, and mAP[0.5–0.95] respectively, showing the best performance among the evaluated methods. The model was tested using unseen images and video clips, and the ensemble approach consistently improved detection results compared to individual models.