This study addresses the pressing need for enhanced security measures in Peru, where recent data from an INEI survey indicates that 25% of residents in major cities have fallen victim to crime. We present an innovative approach to bolstering public safety by employing Transformer neural networks for real-time firearm detection in surveillance cameras. Diverging from conventional methods such as YOLO, our research capitalizes on the capabilities of the Real-Time Detection Transformer (RT-DETR), which utilizes a Hybrid Encoder to facilitate real-time object detection without compromising accuracy. Our model, evaluated on Nvidia Jetson AGX Xavier, achieved a remarkable F1 score of 99.52 % at 38 frames per second, affirming the feasibility of deploying Transformer models on low-power embedded devices by implementing in CUDA. Our findings indicate that Transformer models have the potential to significantly enhance real-time threat detection and fortify urban security infrastructure, presenting a proactive solution to combat the rising challenge of firearm-related crimes. Source code available at https://github.com/labt1/GunDetection-RTDETR.

Automating the detection of weapons from video surveillance images is a difficult task due to: lighting, focus, resolution, among others. Solving this problem would be very useful for citizen security purposes. In this sense, this research work trains a weapon detection system based on YOLOv5 (You Only Look Once) for different data sources, reaching an accuracy of 98.56 % in video surveillance images, performing Real-Time inferences reaching 33 fps on Nvidia's Jetson AGX Xavier which is a good result compared to other existing research in the state of the art.