Alejandro Mesa Gomez
AI-powered high-performance unsafe driving detection on edge devices using Deep Learning, augmented by vehicle CAN bus signals.
Rel. Andrea Bottino. Politecnico di Torino, Master of science program in Data Science And Engineering, 2026
Abstract
Driver distraction and unsafe in-cabin activities continue to constitute significant road-safety challenges, especially for vision-based monitoring systems that must function with limited computing resources and whose safety relevance may depend on the vehicle state. This study examines an edge-oriented deep learning pipeline for driver activity recognition, featuring a modular architecture that separates visual perception from context-aware safety decision-making. The perception component was developed using the Driver Monitoring Dataset, reorganized into 6,287 video segments from 29 subjects across three behavior categories: Normal Driving, Reaching Activity, and Potentially Distracting Activity. To provide a realistic estimate of generalization, the experimental protocol enforced subject-independent training, validation, and test partitions and evaluated transfer learning strategies using a MobileNetV3-Large backbone.
The experimental analysis identified methodological issues that can markedly distort performance estimates in video-based driver monitoring
Relators
Academic year
Publication type
Number of Pages
Additional Information
Course of studies
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modify record (reserved for operators) |
