Abstract
Driver fatigue is one of the most common causes of road accidents, which means that there is a great need for robust and adaptive monitoring systems. Current models of fatigue detection suffer from domain-specific limitations in generalizing across diverse environments, sensor variability, and individual differences. Moreover, they are not resilient to real-time sensor quality issues or missing data, which limits their practical applicability. To overcome the aforementioned challenges, we propose a holistic Dynamic Cross-Domain Transfer Learning framework for fatigue monitoring application using multi-modal sensor data fusion. There are four innovations involved with this framework. Firstly, the domain adversarial neural network in EEG, ECG, and video inputs ensures cross-domain invariance of features. The gap of adaptation at the domain goes below 5%, while there is an improvement of the cross-domain accuracy to as high as 15% from 10%. The ASF-Transformer uses adaptive cross-modal attention for fusing heterogeneous sensor data effectively. Accuracy improves by 5-8% and remains robust under modality dropout conditions. Third, the GMSN dynamically evaluates sensor quality and selectively enables modalities to mitigate performance drops to < 5% even with noisy or missing inputs in process. Fourth, Online Personalized Fine-Tuning (OPFT) allows for real-time adaptation of the model to individual drivers, achieving an improvement in accuracy by 5-7% within 2 h with a latency of < 50ms. Thorough evaluations show that the framework can achieve 85-90% accuracy on target domains while maintaining robustness under 20% sensor dropout. Addressing the issue of domain variability, sensor quality, and personalization, this work has improved the reliability, adaptability, and real-time feasibility of fatigue monitoring systems to provide significant advancements for driver safety in dynamic real-world environments.