Comparative Analysis of Vision Transformers and CNN Models for Driver Fatigue Classification

This study provides a comprehensive evaluation of Convolutional Neural Network (CNN) and Vision Transformer (ViT) models for driver fatigue classification, a critical issue in road safety. Using a custom driving behavior dataset, state-of-the-art CNN and ViT architectures, including VGG16, EfficientNet, MobileNet, Inception, DenseNet, ResNet, ViT, and Swin Transformer, were analyzed in this study to determine the best model for practical driver fatigue monitoring systems. Performance metrics such as accuracy, F1-score, training time, inference time, and frames per second (fps) were assessed across different hardware platforms, including a high-performance workstation, Raspberry Pi 5, and a desktop with a Graphic Processing Unit (GPU). Results demonstrate that CNN models, particularly VGG16, achieve the best balance between accuracy and efficiency, with an F1-score of 0.97 and 77.00 fps on a desktop. On the other hand, Swin V2S outperforms all models in terms of accuracy, achieving an F1-score of 0.99 and 61.18 fps on a GPU, although it exhibits limited efficiency on embedded systems. This study significantly contributes by providing practical recommendations for selecting models based on performance needs and hardware constraints, highlighting the suitability of ViTs for high-computation environments. The findings support the development of more efficient driver fatigue monitoring systems, offering practical implications for enhancing road safety and reducing traffic accidents. ABSTRAK: Kajian ini merupakan penilaian komprehensif terhadap model Konvolusi Rangkaian Neural (CNN) dan Transformer Penglihatan (ViT) bagi pengelasan keletihan pemandu, iaitu satu isu kritikal dalam keselamatan jalan raya. Menggunakan set data tingkah laku pemanduan tersuai, seni bina terkini CNN dan ViT, termasuk VGG16, EfficientNet, MobileNet, Inception, DenseNet, ResNet, ViT dan Transformer Swin dianalisa dalam kajian ini bagi menentukan model terbaik bagi sistem pemantauan keletihan pemandu yang praktikal. Metrik prestasi seperti ketepatan, skor F1, masa latihan, masa inferens, dan bingkai sesaat (fps) telah dinilai merentasi pelbagai platfom perkakasan, termasuk stesen kerja berprestasi tinggi, Raspberry Pi 5, dan komputer meja dengan Unit Pemprosesan Grafik (GPU). Dapatan kajian menunjukkan bahawa model CNN, khususnya VGG16, mencapai keseimbangan terbaik antara ketepatan dan kecekapan, dengan skor F1 sebanyak 0.97 dan 77.00 fps pada komputer meja. Sebaliknya, Swin V2S mengatasi semua model dari segi ketepatan, mencapai skor F1 sebanyak 0.99 dan 61.18 fps pada GPU, walaupun menunjukkan kecekapan yang terhad pada sistem terbenam. Kajian ini memberikan sumbangan yang signifikan dengan menyediakan cadangan praktikal bagi pemilihan model berdasarkan keperluan prestasi dan kekangan perkakasan, serta menonjolkan kesesuaian ViT bagi persekitaran berkomputasi tinggi. Penemuan ini menyokong pembangunan sistem pemantauan keletihan pemandu yang lebih cekap, dengan implikasi praktikal bagi meningkatkan keselamatan jalan raya dan mengurangkan kemalangan.