Face Detection and Recognition in CCTV Security: A Comparative Study of YOLOv5 and YOLOv8

Abstract

Security systems traditionally rely on CCTV for monitoring spaces accessible only to authorized personnel, yet they struggle with face detection and recognition at distances beyond a few meters. This limitation hampers their effectiveness in enhancing room security. This study addresses this challenge by developing a remote facial recognition system utilizing CCTV cameras to identify faces from 1-3 meters away. We employed YOLOv5 and YOLOv8 algorithms, testing pre-trained models of varying sizes (M and X) to improve detection accuracy. The training phase involved 200 epochs with a batch size of 32, yielding mean Average Precision (mAP) scores of 82.7%, 83%, 85%, and 85.2% for YOLOv5m, YOLOv5x, YOLOv8m, and YOLOv8x, respectively. Offline evaluations demonstrated average accuracy rates of 94%, 95%, 90%, and 91%. Online testing, conducted under varying conditions with 1-3 faces visible, showed YOLOv5x achieving an accuracy of 87.8%, compared to 80.9% for YOLOv8x. The results indicate that while single-face recognition is quick and accurate, performance declines with multiple faces in view. This research offers a promising solution to enhance room security through effective facial recognition at a distance, highlighting the potential of improved surveillance technology in secure environments.