Eye Region Modeling Using Statistical Shape Models for Driver Attention Assessment in Advanced Driver Assistance System

Driver inattention continues to be among the top reasons behind road traffic accidents, hence the need for effective and intelligible driver monitoring technologies in the development of Advanced Driver Assistance Systems (ADAS). Despite their precision, current deep learning methods for gaze detection suffer from a lack of intelligibility due to their black-box nature, limiting their transparency and trustworthiness in safety-critical applications. This study proposes an explainable Statistical Shape Model (SSM)-based framework for driver gaze and attention estimation using eye region geometric deformation. The framework constructs a Point Distribution Model from 55 eye-region landmarks across 20,000 samples, applying Generalized Procrustes Alignment (GPA) followed by Principal Component Analysis (PCA) to derive 10 interpretable shape parameters that collectively explain 98.5% of total shape variation. These parameters are physically mapped to horizontal and vertical iris displacement, eyelid aperture, and eye width, enabling transparent estimation of gaze direction. Yaw and pitch gaze angles are recovered through linear regression on shape parameters, achieving a Mean Absolute Error (MAE) of 4.8612° and 4.6823°, and coefficients of determination (R²) of 0.9510 and 0.9310, respectively. The SSM framework is further benchmarked against ResNet50 and MobileNetV2, demonstrating competitive performance with the added benefit of full interpretability. Results confirm that eye-region geometric modeling alone is sufficient to reliably recover driver gaze and detect distraction under real-world sensing constraints, offering a practical and explainable alternative to black-box deep learning methods for ADAS deployment.