Evaluating the Performance of YOLO Architectures for Effective Gun and Knife Detection

Abstract

In recent years, the rise in mass shootings in Thailand has highlighted the need for more comprehensive and cost-effective security solutions. One approach is using artificial intelligence to assist human security personnel, particularly for weapon detection through security cameras. Although advancements in deep learning and computer vision have made it possible to deploy such systems on edge computing devices, real-time weapon detection still faces challenges like accuracy and latency. This study addresses the gap in weapon detection research specific to Thailand by utilizing a dataset featuring local environments and weapons, which differ from those in existing datasets. We compare the performance of YOLO versions 5 through 8, focusing on their mean average precision (mAP) in detecting guns and knives. Since each YOLO version is developed by different research teams and may perform differently under specific conditions, our evaluation considers these variations. The findings indicate that YOLOv8 achieves the highest mAP, with scores of 0.874 on the validation set and 0.848 on the test set, demonstrating its effectiveness in the Thai context.