Estimación Geométrica del Cabeceo en Imágenes Térmicas Mediante Landmarks Faciales y una Red Neuronal Convolucional Multitarea

Resumen

Monitoring driver visual behavior provides valuable information about the driver’s level of attention and plays an important role in developing intelligent driver assistance systems. This paper presents a multitask Deep learning model that uses thermal facial images to estimate facial landmarks and classify overall head orientation (frontal and lateral) simultaneously. The predicted outputs are then used to compute two sets of geometric descriptors for estimating head pitch. The proposed model was evaluated on the Charlotte-Thermal Face dataset, achieving a mean absolute error (MAE) of approximately 9 pixels for facial landmark localization and 87% classification accuracy for head orientation estimation. Furthermore, the proposed descriptors exhibited consistent behavior across different head inclinations, providing complementary geometric information that can be incorporated as an additional feature in driver monitoring systems.