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

Area of Science:

• Computer Vision
• Road Safety
• Machine Learning

Background:

• Driver distraction is a critical road safety issue.
• Effective in-vehicle monitoring systems are essential for mitigating risks.
• Existing systems often struggle with unseen driver generalization and real-time performance.

Purpose of the Study:

• To develop a reproducible evaluation framework for driver distraction detection.
• To assess lightweight models for unseen-driver generalization, explainability, and embedded inference.
• To identify optimal models balancing performance and deployment feasibility.

Main Methods:

• Developed a driver-disjoint evaluation framework for robust generalization testing.
• Benchmarked lightweight models including MobileNetV3-Large on the State Farm Distracted Driver Detection dataset.
• Utilized Explainable AI (XAI) for model interpretability and TensorRT for embedded optimization.

Main Results:

• MobileNetV3-Large achieved 88.92% test accuracy and 97.88% Top-3 accuracy on unseen drivers.
• Explainable AI analysis confirmed focus on behavior-relevant regions (hands, face, phone).
• TensorRT optimization on Jetson Orin Nano Super yielded 212.94 FPS with 4.67 ms latency.

Conclusions:

• MobileNetV3-Large offers a practical solution for real-time driver distraction detection.
• The proposed framework enables comprehensive evaluation of driver monitoring systems.
• Achieved a balance between generalization, interpretability, and efficient embedded deployment.