Enhancing Real-Time Detection Transformer (RT-DETR) for Handgun Detection on Nvidia Jetson

Abstract

Recent studies have highlighted the rise in violence and criminal activities primarily
involving firearms. In response to the growing demand for effective firearm detec-
tion systems in public safety applications, this paper explores advancements in real-
time object detection using Transformer-based models. Building on the RT-DETR
architecture and its latest version, RT-DETR v2, we introduce improvements such as
the Bidirectional Feature Pyramid Network (BiFPN) for enhanced small object de-
tection and dynamic batch processing to maximize computational efficiency and re-
source utilization on edge devices like the Nvidia Jetson AGX Xavier for efficient real-
time deployment. We also compare the model with state-of-the-art alternatives such
as YOLOv10, demonstrating the superiority of Transformer models in terms of accu-
racy and performance. For the comparative study, we used a benchmark proposing
three datasets with challenging conditions. Code and trained models are available at
https://github.com/labt1/GunDetection-RTDETR.