Performance Evaluation of YOLOv3, YOLOv4 and YOLOv5 for Real-Time Human Detection

Abstract

The main concern of human detection using computer vision is to correctly identify people in an image and video. Human detection has been a topic of intensive study over the last decade. YOLO being single stage algorithms happen to offer better speed than two stage algorithms hence making them a better choice for real time object detection. This strategy has the benefit of offering a comprehensive study of contemporary human detection techniques as well as a manual for selecting the best ones for actual applications. In addition, Real-time human detection and occlusion issues are also looked at. In this paper, experimentation is done on real time image to verify the performance of different models of YOLO family i.e YOLOv3, YOLOv4 and YOLOv5. The experiment shows that YOLOv5 is best performer in terms of mAP with precision of 0.84 while YOLO v3 is the fastest but with a slightly less precision of 0.71. The mAP of the three algorithms were 0.86, 0.89 and 0.91 respectively.