Deep Learning-Based Real-Time Driver's Hands-On Detection: A Lightweight Time Series Approach Using CAN Data

초록

Advanced Driver Assistance Systems (ADAS) significantly enhance vehicle safety and driving convenience by integrating sensors, cameras, and machine learning algorithms, with Hands-On Detection (HOD) being critical for ensuring driver readiness for manual control. Existing HOD methods face limitations like false positives, environmental sensitivity, and high computational demands. In this study, we propose a novel deep learning framework leveraging Controller Area Network (CAN) data treated as a time series classification problem, introducing the construction of a new CAN-based HOD dataset from realistic driving scenarios and the design of an efficient deep learning architecture for real-time deployment. Specifically, we present the Efficient HOD Network (EHODNet), whose core design is centered on the proposed TCSC module, enabling effective multiscale temporal-channel feature modeling with a hierarchical network structure. This approach provides an efficient and effective representation of HOD characteristics, bridging the gap between existing time-series classification methods by improving the trade-off between computational complexity and performance. Furthermore, we preprocess the CAN data to reduce noise and enhance robustness, crucial for accurate feature extraction. Experimental results confirm that EHODNet offers superior accuracy and computational efficiency, providing a robust, real-time solution suitable for practical ADAS applications.

키워드