Weapon detection system for surveillance and security

Abstract

Weapon detection is a critical and serious topic in terms of public security and safety, however, it's a difficult and time-consuming operation. Due to the increase in demand for security, safety, and personal property protection, the requirement for deploying video surveillance systems capable of recognizing and interpreting scenes and anomalous occurrences plays an important role in intelligence monitoring. In certain regions of the globe, mass shootings and gun violence are on the increase. Timely detection of the presence of a gun is critical to prevent loss of life and property. Several object detection models are available, which struggle to recognize firearms due to their unique size and form, as well as the varied colors of the background. In this study, we proposed a state-of-the-art system based on a deep learning model YOLO V5 for weapon detection that will be sufficiently resilient in terms of affine, rotation, occlusion, and size. We evaluated the performance of our system on a publicly available dataset and achieved the F1-score of 95.43%. For the detection and segmentation of firearms, we implemented Instance segmentation or Pixel level segmentation which employed Mask-RCNN. The system achieved the detection accuracy (DC) of 90.66% and 88.74% Mean intersection over union (mIoU). The purposed methodology combined different data augmentation and preprocessing methods to improve the accuracy of the proposed weapon detection system.