Detection of suspicious facial objects in neutral ATMs using deep learning architectures based on YOLOV8 and Faster R-CNN

This study presents an automatic detection system for suspicious facial objects in neutral automated teller machines (ATMs), using and comparing the deep learning architectures YOLOv8 and Faster R-CNN. A dataset was built from real ATM surveillance videos and complementary images, from which frames were extracted and annotated with masks, hats, and glasses. Both models were trained under the same preprocessing pipeline and evaluated using standard object detection metrics such as precision, recall, F1-score, Intersection over Union (IoU), and mean Average Precision (mAP), in order to analyze their performance in real surveillance conditions. The results show that YOLOv8 achieves higher precision, reducing the generation of false positives, while Faster R-CNN demonstrates higher recall and superior mAP@0.5 values in several classes, indicating greater sensitivity to partially visible objects. A decision-making logic was also integrated to automatically classify each scene as NORMAL or SUSPECT, based on the combined presence of facial-occluding elements. The implementation demonstrates that computer vision systems can complement security mechanisms in neutral ATMs by providing early detection of potential risks and enabling real-time remote monitoring.