Analytical Study of YOLO and Its Various Versions in Crowd Counting

Abstract

Crowd counting is one of the main concerns of crowd analysis. Estimating density map and crowd count in crowd videos and images has a large application area such as traffic monitoring, surveillance, crowd anomalies, congestion, public safety, urbanization, planning and development, etc. There are many difficulties in crowd counting, such as occlusion, inter and intra scene deviations in perception and size. Nonetheless, in recent years, crowd count analysis has improved from previous approaches typically restricted to minor changes in crowd density and move up to recent state-of-the-art systems, which can successfully perform in a broad variety of circumstances. The recent success of crowd counting methods can be credited mostly to the deep learning and different datasets published. In this paper, a CNN-based technique named You Only Look Once (YOLO), and its various versions have been studied, and its latest version, YOLOv5, is analyzed in the crowd counting application. This technique is studied on three benchmark datasets with different crowd densities. It is being observed that YOLOv5 gives favorable results in crowd counting applications with density ranges from low to medium but not in a very dense crowd.