Comparative Performance Analysis of YOLOv4 and YOLOv5 Algorithms on Dangerous Objects

Abstract

It is very important that the legal order is carried out without interruption and that people can live without fear. This can only occur in a high security environment. For this reason, today’s security mechanisms are constantly evolving in order for life in society to be able to live safely. In particular, many locations have surveillance systems with cameras in order to ensure environmental safety in recent years. Images obtained from these cameras can be analyzed using computer vision techniques to create auxiliary systems. One of the sub-branches of artificial intelligence, the visual technique is inspired by the ability of people to understand, identify and detect objects, and is the solution to many problems today, thanks to the development of hardware components and data storage systems. The main purpose of the computer vision technique, which provides solutions to problems such as pedestrian detection, autonomous driving systems etc., is to enable visual data to be understood, identified and classified with artificial intelligence techniques. Many deep learning-based algorithms are used to perform these operations. The operating principles and performance of these different algorithms vary. Especially among the latest versions of YOLO, one of the popular object detection algorithms performance comparison varies from various sources. Based on this situation, within the scope of the study, performance comparison was performed between YOLOv4 and YOLOv5 models. To the models are tasked with identifying the gun and the person carrying it that could threaten security when used by unauthorized persons. A data set of guns and people were used to make this test possible. Since the quality and labeling of the dataset to be used is known to be very important to the success of the model, the data is collected by us and tagged one by one. The obtained data were used in YOLOv4 and YOLOv5 models with the same hyper-parameters, and the results produced by both models were compared and analyzed. According to the results obtained, we confirm that these algorithms can be used to detect dangerous objects, and we report the performance values generated for this application with details. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.