YOLO-Based Anomaly Detection and Alert System for ATM Security

- Automated teller machines (ATMs) are usually watched by CCTV cameras that only record, so suspicious behaviour is normally discovered after the damage is done. This paper presents a low-cost anomaly detection and alert system for ATM kiosks built on the YOLOv8n object detector. The system looks for two warning signs: a user whose face is hidden by a mask or helmet, or is not visible at all, and a person who enters the restricted service area of the machine, which is observed by a separate overhead camera. To avoid false alarms from momentary occlusion, an anomaly is confirmed only after it persists for four consecutive seconds. Once confirmed, the system saves a time-stamped evidence image and sends a Telegram message with the image attached to the bank's security desk. A prototype was implemented on an ordinary laptop with webcams, without dedicated edge-AI hardware, and was trained on a hand-labelled five-class dataset. Live screenshots from the prototype show that the face-concealment module separates a visible face from a masked one and raises an alert for the latter. The work indicates that real-time monitoring of more than one ATM-specific threat is practical on inexpensive commodity hardware.