Use of Custom Videogame Dataset and YOLO Model for Accurate Handgun Detection in Real-Time Video Security Applications

• Autores: Bazan, Diego, Casanova, Raul, Ugarte, Willy
• Formato: artículo
• Fecha de Publicación: 2024
• Institución: Universidad Peruana de Ciencias Aplicadas
• Repositorio: UPC-Institucional
• Lenguaje: inglés
• OAI Identifier: oai:repositorioacademico.upc.edu.pe:10757/676064
• Enlace del recurso: https://doi.org/10.5220/0012716500003690; http://hdl.handle.net/10757/676064
• Nivel de acceso: acceso abierto
• Materia: Artificial Vision; Criminal Activities; Custom Pistol Video-Game Dataset; Human Limitations; Machine Learning; Real Time Detection; Video Surveillance Systems; YOLOV7; https://purl.org/pe-repo/ocde/ford#3.00.00

Research has shown the ineffectiveness of video surveillance operators in detecting crimes through security cameras, which is a challenge due to their physical limitations. On the other hand, it was shown that computer vision, although promising, faces difficulties in real-time crime detection due to the large amount of data needed to build reliable models. This study presents three key innovations: a gun dataset extracted from the Grand Theft Auto V game, a computer vision model trained on this data, and a video surveillance application that employs the model for automatic gun crime detection. The main challenge was to collect images representing various scenarios and angles to reinforce the computer vision model. The video editor of the Grand Theft Auto V game was used to obtain the necessary images. These images were used to train the model, which was implemented in a desktop application. The results were very promising, as the model demonstrated high accuracy in detecting gun crime in real time. The video surveillance application based on this model was able to automatically identify and alert about criminal situations on security cameras-