The Journal of Science and Engineering Applications (JSEA)

Abstract

In recent years, the global rise in crimes involving weapons has become more pronounced, 
particularly in areas with weak law enforcement or where firearm possession is legal. To address this 
issue, early identification of suspicious behavior related to weapon possession is crucial, enabling law 
enforcement agencies to take swift action. While the human visual system is highly advanced and capable 
of processing images quickly and accurately, prolonged observation of similar visual data can lead to 
fatigue and reduced attentiveness. Additionally, large-scale surveillance systems with numerous cameras 
require extensive monitoring teams, which drives up operational costs. Several automatic weapon 
detection solutions based on computer vision have been proposed, but their performance remains limited 
in challenging environments. A systematic review of current literature on deep learning-based weapon 
detection was conducted to assess the methods employed, the characteristics of existing datasets, and the 
key challenges in automatic weapon detection. The most frequently utilized models were Faster R-CNN 
and YOLO architectures, and the integration of realistic images with synthetic data showed improved 
detection accuracy. Major challenges include poor lighting conditions and difficulties in detecting small 
weapons, with the latter being particularly significant. The focus of this review is to examine the methods 
used for detecting and tracking weapons using deep learning techniques.