Edge artificial intelligence for detecting bladed weapons: high-availability architecture with human supervision in the PMPA

Abstract

The operational reality of the Military Police of Pará (PMPA) involving bladed weapons, marked by the seizure of 7,679 units in 2024, demands more efficient surveillance systems to support real-time police action. This article proposes the development of a computer vision system based on artificial intelligence for the automatic detection of such artifacts, architected to operate on edge servers within the PMPA structure, allowing identification with low latency and greater reliability. The system is conceived as a support tool under human supervision, where police action is only requested when potential threats are identified, reducing the need for constant monitoring and, consequently, the cognitive fatigue of operators, a factor that could make detection time-consuming and prone to errors, preserving the reliability of the incident by filtering out false alarms. The methodology involved training a YOLO26l model with a dataset of 2,356 post-sanitization images, focused on real-world scenarios of low light and partial occlusion, employing data balancing strategies and the intensive use of negative images (Hard Negative Mining), representing 28.4% of the total, to reduce false alarms. The results demonstrated an average accuracy of 97.19% (mAP50) with a processing speed of 28.9 FPS. These figures prove the tool's efficiency as a force multiplier, allowing for expanded surveillance capacity without the need to increase personnel.