World Journal of Engineering Research and Technology

Lyndon Bermoy*, Jecelyn Sanchez

ABSTRACT

Driver distraction remains one of the leading causes of vehicular accidents worldwide, necessitating intelligent monitoring systems that can accurately detect inattentive behavior in real time. This study presents the development and evaluation of a vision-based driver distraction detection system that integrates hybrid eye-tracking and 3D head pose estimation to enhance attention recognition accuracy across varying driving conditions. The proposed system fuses Convolutional Neural Network (CNN)-based gaze estimation using MobileNetV3 with geometric head pose estimation derived from 68-point facial landmarks and solved via the Perspective-n-Point (PnP) algorithm. Both modalities were processed through a fusion layer to classify three attention states: Focused, Temporarily Distracted, and Critically Distracted. The hybrid framework was implemented on a Raspberry Pi 5 coupled with a Coral TPU accelerator for real-time inference. A custom dataset consisting of video frames captured under bright daylight, normal indoor, and dim lighting conditions was used for model training and validation. Experimental results revealed that the hybrid model achieved an overall accuracy of 95.2%, outperforming single-modality models—CNN-only (91.5%) and Head Pose-only (88.3%)—by a significant margin. Statistical validation using ANOVA confirmed that differences in model performance were significant (F(2, 27) = 9.45, p < 0.01). The system demonstrated strong resilience to partial occlusions such as eyewear, hand-over-face and variable illumination, maintaining a false alarm rate below 5%. This study concludes that integrating eye-tracking with 3D head pose estimation provides a reliable and efficient approach for driver attention monitoring. The proposed hybrid model provides a scalable foundation for real-time, embedded driver-assistance systems that can enhance road safety by reducing distraction-related incidents.

[Full Text Article] [Download Certificate] https://doi.org/10.5281/zenodo.17750065