Scholarlink

<p>Driver distraction is a major contributor to traffic accidents worldwide, prompting the need for real-time focus detection systems. This study proposes a deep learning-based approach to classify driver attention using eye movement data captured through eye tracking. A new dataset was collected from 15 respondents during simulated driving sessions, from which 32,362 images were extracted and labeled as either focused or unfocused. Three convolutional neural network (CNN) models were developed and evaluated: two based on transfer learning (Inception V3 and MobileNet V2), and one full learning model using architecture optimization via the Taguchi method. The Inception V3 model achieved the highest classification performance, with an average accuracy of 76.78% and an F1 score of 0.76. The custom full learning model achieved 71.88% accuracy with the shortest inference time, while MobileNet V2 yielded the lowest performance. These findings demonstrate that eye-tracking images can serve as effective input for visual attention modeling in driver monitoring systems. The results also highlight a trade-off between accuracy and efficiency, offering valuable insights for real-time applications in intelligent transportation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Driver Focus Detection Based on Eye Tracking Images Using Convolutional Neural Networks: A Comparative Study of Transfer Learning and Custom Architectures

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Haidar Rabbani,
  • Achmad Pratama Rifai,
  • Wangi Pandan Sari,
  • Andri Nasution

摘要

Driver distraction is a major contributor to traffic accidents worldwide, prompting the need for real-time focus detection systems. This study proposes a deep learning-based approach to classify driver attention using eye movement data captured through eye tracking. A new dataset was collected from 15 respondents during simulated driving sessions, from which 32,362 images were extracted and labeled as either focused or unfocused. Three convolutional neural network (CNN) models were developed and evaluated: two based on transfer learning (Inception V3 and MobileNet V2), and one full learning model using architecture optimization via the Taguchi method. The Inception V3 model achieved the highest classification performance, with an average accuracy of 76.78% and an F1 score of 0.76. The custom full learning model achieved 71.88% accuracy with the shortest inference time, while MobileNet V2 yielded the lowest performance. These findings demonstrate that eye-tracking images can serve as effective input for visual attention modeling in driver monitoring systems. The results also highlight a trade-off between accuracy and efficiency, offering valuable insights for real-time applications in intelligent transportation.