Zone Intrusion Detection

Problem

On-site PPE monitoring relies on the human eye. A worker entering a restricted zone without a helmet or vest goes unnoticed if the supervisor happens not to be watching; the violation often surfaces only after an accident.

Approach

I built a dual-model architecture that combines a custom-trained YOLOv8 model with the pre-trained YOLOv8n person-detection model. The custom model was trained for 50 epochs and reached 88% mAP50 accuracy. The user draws square, triangular, and rectangular restricted zones onto the camera view with the mouse; zones can be moved and resized via drag-and-drop. The system flags, in real time, anyone entering those zones without protective equipment.

Result

On a violation, an audible alarm and an on-screen warning kick in, and a photo notification is sent through the Telegram Bot API. For every violation, the system generates an automatic Turkish incident report using the Groq API (LLaMA 3.1).