Sentry: Efficient Real-Time Weapon Detection on Edge-Mobile Platforms

Alarcón S.; Elsner J.; Navarrete E. · 4th Cognitive Models and Artificial Intelligence Conference, AICCONF 2026 - Proceedings (2026) · DOI: 10.1109/aicconf69182.2026.11600643

Resumen

Urban public safety in Lima, Peru faces persistent challenges related to weapon enabled crime, where timely detection and rapid response are critical. This paper presents Sentry, an Edge IoT mobile platform for real time weapon detection that integrates on device computer vision, geolocation, and push notifications to support municipal patrol operations. The system follows a hybrid edge cloud architecture. Detection is executed locally on mobile devices to reduce latency and preserve privacy, while cloud services provide authentication, event logging, and notification delivery. To train and evaluate the detection model, we curated and annotated a dataset of 4100 images containing firearms, knives, and non weapon contexts. Images were resized to 640 by 640 and augmented using flipping, rotation, cropping, and zooming to improve robustness to real world variations in viewpoint, scale, and illumination. The dataset was split into 70 percent training, 20 percent validation, and 10 percent test subsets. On the held out test set, the proposed YOLOv8 based detector achieved mean average precision at intersection over union 0.50 of 0.909 and mean average precision averaged from 0.50 to 0.95 of 0.655, with precision 0.928, recall 0.878, and F1 score 0.902. The proposed design supports real time alerting under resource constraints and offers a practical, privacy aware alternative to fully centralized surveillance pipelines in urban environments.

Instituciones: Universidad Peruana de Ciencias Aplicadas