Weapons detection in video surveillance systems using deep neural networks

Abstract

The purpose of this paper is to conduct an empirical assessment of the usefulness of selected deep neural network architectures in detection of weapons in CCTV images. The analysis compa- red single-stage YOLOv5 detectors and two-stage Faster R-CNN detectors, each trained under identical conditions on a limited dataset. The experiment used a dataset from the Kaggle platform comprising 4,000 images of firearms and knives. The models were trained with transfer learning on COCO pre-trained weights, and their performance was assessed in terms of mAP, precision, re- call and frames per second (FPS). The results obtained indicate that the YOLO family offers a favo- urable trade-off between accuracy and speed, enabling video stream processing in near real time