A lightweight deep learning model for real-time in-vehicle driver distraction detection with low-latency inference

Driver distraction is a major road-safety concern that requires reliable and efficient in-vehicle monitoring systems. The main contribution of this work is a reproducible driver-disjoint and deployment-oriented evaluation framework that jointly examines unseen-driver generalization, lightweight model benchmarking, explainability, calibration, and embedded inference. Experiments on the State Farm Distracted Driver Detection dataset show that MobileNetV3-Large provides the best trade-off among the evaluated lightweight models, achieving 88.92% test accuracy, 89.02% balanced accuracy, 88.07% macro-F1, and 97.88% Top-3 accuracy on unseen drivers. Explainable AI analysis indicates that the model mainly focuses on behavior-relevant regions, including the hands, face, phone area, steering wheel, and upper-body posture. For embedded deployment, TensorRT optimization on the Jetson Orin Nano Super achieved 212.94 FPS with 4.67 ms end-to-end latency in FP16 mode. These results demonstrate a practical balance between unseen-driver generalization, interpretability, and real-time embedded inference.