Benchmark Comparative Analysis of YOLO Architectures for Weapon Detection with Explainability

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Firearms and other weapons pose significant public safety risks due to their widespread accessibility. This research paper investigates automated detection of dangerous objects, specifically guns and persons, to support early intervention in high-risk situations. We present a comprehensive benchmark evaluation using the YouTube Gun Detection Dataset (YouTube-GDD), which contains 5,000 high-definition images with 16,064 gun instances and 9,046 person instances across two classes. Ten You Only Look Once (YOLO) model variants (v5, v8, v10, v11, and 26) in both nano and small configurations were evaluated using mAP50, mAP50-95, inference speed, and model size as key performance metrics. Furthermore, explainable AI methods using occlusion-based saliency maps were applied to the highest performing models to provide interpretable visualisations of model attention regions and increase detection reliability. YOLOv8s achieved the highest mAP50 of 0.876, while YOLO26s reached the best mAP50-95 of 0.768. The fastest model was YOLOv5n at 1018.4 FPS using an NVIDIA GeForce RTX 4080 Laptop GPU with 12 GB VRAM. This work establishes the first comprehensive benchmark for this dataset and represents the first application of explainability methods to dangerous object detection.