Deep Learning-Based Real-Time Weapon Detection System

Abstract

In recent years, the rate of gun violence has risen at a rapid pace. Most current security systems rely on human personnel to monitor lobbies and halls constantly. With the advancement of machine learning and, specifically, deep learning techniques, future closed-circuit TV (CCTV) and security systems should be able to detect threats and act upon this detection when needed. This paper presents a security system architecture that uses deep learning and image-processing techniques for real-time weapon detection. The system relies on processing a video feed to detect people carrying different types of weapons by periodically capturing images from the video feed. These images are fed to a convolutional neural network (CNN). The CNN then decides if the image contains a threat or not. If it is a threat, it would alert the security guards on a mobile application and send them an image of the situation. The system was tested and achieved a testing accuracy of 92.5%. Also, it was able to complete the detection in as fast as 1.6 seconds.