Deep Learning for Driver Monitoring: A Review of Methods, Challenges, and Future Directions
Year: 2026
Type: article
Abstract: Driver monitoring has become an important application of deep learning in intelligent transportation systems because driver distraction, drowsiness, and unsafe driving behaviour can increase road-safety risks. Recent studies have investigated convolutional neural networks, vision transformers, pose estimation, temporal learning, and lightweight architectures to recognize driver activities under different conditions. However, these methods often concentrate on an individual monitoring task or evaluate related tasks using different datasets and experimental protocols. This survey reviews ten IEEE journal papers on driver distraction, activity recognition, drowsiness detection and fatigue analysis, nine of which were published in 2024 or 2025. The reviewed studies are compared in terms of input data, model architecture, learning strategy, reported performance, computational efficiency, and limitations. Most of the vision-based methods report accuracy between 96% and 99.8% on their own test data, but one cross-dataset test shows accuracy falling to 73.88%, and none of the reviewed studies predicts distraction, drowsiness and seatbelt use together. The comparison highlights the potential of combining spatial feature extraction with temporal information while also identifying challenges associated with differences between datasets, driver diversity, lighting conditions, generalization, and real-time deployment. Building on these findings, the survey outlines a multi-task monitoring framework with a shared feature extractor and separate prediction heads for distraction, drowsiness, and seatbelt-use recognition. The proposed direction aims to reduce duplicated computation while maintaining task-specific predictions. Future evaluation should compare the multi-task model with independent single-task models using subject-independent testing, class-wise performance measures, inference latency, and cross-dataset evaluation. The survey provides a structured basis for developing more efficient and comprehensive driver-monitoring systems.
Authors: Yadav, Rohit, Chawan, Pramila M.
Language: English
Subfield: Human Factors and Ergonomics
Field: Social Sciences
Domain: Social Sciences
Keywords: driver monitoring, driver distraction detection, drowsiness detection, fatigue analysis, driver activity recognition, multi-task learning, shared feature extractor, cross-dataset generalization, real-time deployment, lightweight architecture, vision transformer, convolutional neural networks, pose estimation, temporal learning, road safety
Open Access status: hybrid
OpenAlex ID: W7225134209
