Rede Neural convolucional aplicada no monitoramento por câmeras de segurança

Abstract

As security is a primary factor for a good quality of life and considering the current situation in our society, the possibility of developing more efficient tools to deal with insecurity is evident. One of the main measures to improve security is video monitoring, taking advantage of advances in software, such as object detectors, and hardware, with the development and dissemination of better video cameras and graphics cards, necessary for high processing of parallel neural networks. Therefore, this project proposes the development of a system capable of identifying people carrying weapons through closed-circuit television images. Based on a research it was concluded that the state of the art in object detection is the YOLO series. There are several studies on the use of neural networks with the same purpose of this project, but it differs from the others because it uses YOLO version 5. The database is composed of images of pistols and revolvers, being created a unique class called weapon. Network performance is measured through its results of precision 97.58%, recall 91.75% and mAP (mean average precision) 94.57%. This performance was achieved using a YOLOv5 small model and a dataset with 4369 images with own images.