Real-Time Surveillance System based on Facial Recognition using YOLOv5

Facial recognition using deep learning techniques is now a rapidly growing and widely applied aspect of real-time surveillance systems with broad range of applications in every field. Recognizing multiple faces in real-time is very challenging due to adverse environmental conditions and occlusion effects. YOLOv5 is the current state-of-the-art algorithm for real-time facial recognition with very limited experimental analysis. In this paper YOLOv5 has been trained from scratch and tested on FDDB and customized dataset from real-time video feed. Experiments show 87% accuracy on FDDB while 94% accuracy on the customized dataset. The paper also presents comparative analysis of the results with the previous versions of YOLOv5 (YOLOv3 and YOLOv4). The algorithm is also tested on real-time environment and has the capability to detect multiple faces with maximum accuracy.