Automatic Weapon and Knife Detection System on Security Cameras: Comparative YOLO Models

Abstract

Security camera systems are generally used to create safe environment conditions in social and public areas. In today’s security camera systems, artificial intelligence-based computer vision has started to be used instead of human vision. In this study, it is aimed to see and detect images of guns and knives that may be in social and public spaces with computer vision. As a method and model in the study, image processing technology and YOLO algorithms, which are generally known to have very successful results in the literature, were used. Among the YOLO algorithms, YOLOv4, YOLOv5, YOLOR and YOLOX models were used. 5078 images were used as a data set in the study, and 3000 of these images consist of images of weapons and 2078 of knife images. In order to ensure reliability for the experimental study to be obtained from the images, attention has been paid to the fact that the selectivity of the images is difficult. Comparative experimental studies of YOLO algorithms have been carried out and their results have been published. The most successful result in detecting weapons and knives in the images was obtained in the YOLOR model with a map@0,5 value of 97.6%.