Vision-based driver distraction classification: a comprehensive survey of deep learning methods, evaluation protocols, and deployment challenges

One of the significant causes of road traffic accidents is driver distraction, which makes driver behavior classification very crucial for Advanced Driver Assistance Systems (ADAS) and driver monitoring systems. A literature analysis suggests that recent vision-based approaches using the SFD3 benchmark dataset tend to achieve high classification accuracy (97.89–99.80%) with lower interquartile range results when the images are split randomly. These high values are typically due to evaluation within the dataset and to splitting data that is not subject-independent. This can lead to image-level and subject-level data leakage, and overly optimistic performance estimates. By contrast, other evaluation methods such as subject-independent or subject-wise cross-dataset results tend to be significantly lower, generalizing to an unseen driver or new datasets. Unlike conventional surveys that mainly categorize architectures and datasets, this survey critically examines the reliability and comparability of reported results. It includes traditional machine learning, Convolutional Neural Networks (CNNs), transfer learning, attention-based models, recurrent networks, Vision Transformers (ViTs), ensembles, and hybrid models. The survey also covers Public and Authors’ datasets, Data preprocessing, Data augmentation, Dataset splitting, Cross-validation, Evaluation metrics, Computational complexity, Explainable Artificial Intelligence (XAI), and embedded deployment. The approaches are systematically compared regarding architecture, datasets, evaluation protocols, predictive performance, generalization, computational requirements, and deployment feasibility. The survey reveals research gaps related to inconsistent evaluation practices, limited subject-independent and cross-dataset validation, computational constraints, and explainability. There are several future directions such as standardized evaluation protocols, advanced data augmentation, efficient Transformer architectures, multimodal learning, and trustworthy deployment-ready driver monitoring systems.