Comparison YOLOv5 Family for Human Crowd Detection

Abstract

Recent years have seen widespread application of crowd counting and detection technology in areas as varied as urban preventing crime, station crowd statistics, and people flow studies. However, getting accurate placements and improving audience counting performance in dense scenes still has challenges, and it pays to devote a lot of effort to it. In this paper, crowd counting models are proposed based on the YOLOv5 algorithm, and four YOLOv5 models (YOLOv5l, YOLOv5m, YOLOv5s, YOLOv5x) were built for the purpose of comparing the models and increasing the accuracy of crowd identification as each model contains certain characteristics such as Filter sizes. Each model was trained on a human dataset (indoor and outdoor) for the purpose of comparing the results of each model and showing which model reaches higher accuracy in detecting people. Through this study and practical experiments conducted on each model, it was found that the best model is YOLOv5x, and YOLOv5l, where the accuracy of detecting humans reached more than 96%, while YOLOv5s reached more than 92%, and YOLOv5m reached the lowest accuracy, which is 91%.