Employing YOLOv4 in a Vehicle Counting System Using Random Dataset

A vehicle counting system (VCS) is one technology that can help the Intelligent Transportation System (ITS) achieve its goals of providing safe and efficient road and transportation infrastructure. Some challenges arise when dealing with CCTV objects, such as an object being blocked by other objects or scanning objects being too small, resulting in the object not being detected. This paper implements YOLOv4, with the number of vehicles based on five classes: car, bus, truck, bicycle, and motorcycle. To demonstrate the feasibility system and adequacy, which implement in three stages: input the videos or images into a system using pre-trained YOLOv4, get the bounding box, and compare with the ground truth to the pre-trained model using the MS COCO dataset. The result of this research is that accuracy of 85.61% is obtained based on the results of the two-test media of Image and Video, which were broken down into frame by frame to get accuracy by frame. For image detection, an accuracy of 88% is obtained; for video detection, an accuracy of 83.22% is obtained.