Evaluation of Convolutional Neural Network Architectures for Detecting Drowsiness in Drivers

• Autores: Hurtado Delgado, Bryan, Oscco Guillen, Marycielo Xiomara
• Formato: tesis de grado
• Fecha de Publicación: 2026
• Institución: Universidad Nacional Micaela Bastidas de Apurímac
• Repositorio: UNAMBA-Institucional
• Lenguaje: español
• OAI Identifier: oai:null:20.500.14195/1562
• Enlace del recurso: https://dx.doi.org/10.14569/IJACSA.2025.0160217; https://hdl.handle.net/20.500.14195/1562
• Nivel de acceso: acceso abierto
• Materia: Architectures; Detection; Drowsiness; Neural networks; https://purl.org/pe-repo/ocde/ford#2.02.04

Drowsiness in drivers is a condition that can manifest itself at any time, representing a constant challenge for road safety, especially in a context where artificial intelligence technologies are increasingly present in driver assistance systems. This paper presents a comparative evaluation of convolutional neural network (CNN) architectures for drowsiness detection, focusing on the identification of signals such as eye state and yawning. The research was of an applied type with a descriptive level, comparing the performance of LeNet, DenseNet121, InceptionV3 and MobileNet under challenging conditions, such as lighting and motion variations. A non-experimental design was used, with two datasets: a public dataset from Kaggle that included images classified into two categories (yawn and no yawn) and another created specifically for this study, which included images classified into three main categories (eyes open, eyes closed and undetected). The results indicated that, although all architectures performed well in controlled conditions, MobileNet stood out as the most accurate and consistent in challenging scenarios. DenseNet121 also showed good performance, while LeNet was effective in eye-state detection. This study provided a comprehensive assessment of the capabilities and limitations of CNNs for applications in drowsiness monitoring systems, and suggested future directions for improving accuracy in more challenging environments.