Autora: Paola Natalia Cañas Rodríguez

Tesis: Driving scene understandig and risk estimation using in-cabin monitoring systems and exterior information

Dirección: Igor Rodríguez / Marcos Nieto

Día: 20 de julio de 2026
Hora: 11:30h
Lugar: sala Ada Lovelace (Facultad de Informática)

Abstract:

"As the automotive industry moves toward higher levels of automation, safe operation still depends on the vehicle’s ability to understand both the driver’s state and the surrounding environment. This thesis addresses the design of robust and trustworthy Driver Monitor- ing Systems (DMS) by covering the entire lifecycle: from data creation and algorithmic development to ethical and regulatory compliance.

First, the work tackles the data bottleneck through the collaboration on the creation of the Driver Monitoring Dataset (DMD), a large-scale synchronized multimodal and multi- sensor benchmark for distraction, drowsiness, and gaze estimation. This is complemented by synthetic data generation and a semi-automatic annotation pipeline. By integrating active learning with human-in-the-loop supervision, this pipeline reduces manual labeling effort while maintaining a label reliability.

Second, the thesis develops integrated vision-based DMS functionalities aligned with Euro NCAP requirements and definitions. The system adopts a sub-set of the defined gaze regions, triggering times of alerts and is designed for robustness against “noise variables” such as age, gender, and extreme lighting variations. The developed DMS include driver identification, gaze estimation, distraction monitoring, and a novel approach to non- functional-state (occlusion) detection using Vision-Language Models (VLMs) to provide semantic understanding of system degradation.

Third, a Dynamic Risk Assessment methodology is proposed that fuses in-cabin awareness with external perception. By combining driver gaze with 3D LiDAR-based scene under- standing within a Local Dynamic Map (LDM), the system enables a human-centered estimation of driving risk. Using reverse parking maneuvers as a use-case, the research demonstrates that risk is a function of both physical proximity and driver situational Awareness."