AN EFFICIENT DEEP LEARNING FRAMEWORK FOR FIREARMS DETECTION IN SECURE SMART CITY INFRASTRUCTURE

Our civilised world cannot tolerate violence in any shape or form. Still, many innocent lives are lost every year due to the pervasiveness of violence in our culture, especially in the current day. One of the conventional means of violence is using a firearm. Deaths caused by firearms are a worldwide problem right now. Both society and law enforcement authorities face this difficulty. Cities and semi-urban regions account for a disproportionate share of these crimes. Today, CCTV surveillance is widely used for both prevention and monitoring by both public and private organisations. Nevertheless, human-based monitoring is error-prone and consummately resource-intensive. However, when it comes to violent behaviours, automated smart surveillance is better suited to scale and dependability. Demonstrating the integration of deep learning-based approaches for weapon detection is the primary goal of this study. In order to identify firearms and people, this study employs a variety of detection methods, including most recently developed EfficientDet-based architectures and Faster Region-Based Convolutional Neural Networks (Faster RCNN). After using post-processing methods such as Weighted Box Fusion, Non-Maximum Weighted, and Non-Maximum Suppression, an ensemble (stacked) approach enhanced the detection performance for both human faces and firearms. In this work, we have presented and compared the findings of many detection methods and ensembles. It aids law enforcement in gathering incident information quickly so they may take preventative actions as soon as possible. It is also possible to use the same method to recognise movies on social media that include firearms. Mean average precisions of 77.02% for mAP0.5, 16.40% for mAP0.75, and 29.73% for mAP[0.500.95] are provided here by the Weighted Box Fusion-based Ensemble Detection Scheme. The outcomes outperform all of the alternatives that were tested. An extensive battery of tests using random test photos and video clips has been conducted on the model. Consistently outperforming main models, the resulting ensemble methods are adequate.