BananaGuard: A Comparative Analysis of AI Models for Firearm Detection in Crowded Environments

Firearm detection in crowded and dynamic environments remains a critical challenge for public safety due to occlusions, complex backgrounds, and high stakes of false positives and negatives. Existing solutions often fail to balance accuracy, computational efficiency, and real-time applicability in resource-constrained settings. To address this issue, we benchmarked state-of-the-art object detection models-YOLOv10, YOLOv11, and RT-DETR-using a rigorous evaluation framework on a diverse dataset of firearm images. Our study demonstrates that YOLOv11 effectively balances real-time performance and accuracy, while RT-DETR excels in nuanced detection tasks, requiring greater computational resources. We emphasize the need to align model selection with operational priorities, offering actionable insights to improve firearm detection systems in real-world scenarios. This work advances the understanding of model trade-offs and provides a foundation for deploying AI-driven