关键词: 方向盘扭矩；脱手检测；多特征融合；深度学习

Abstract: In L2-level autonomous driving assistance systems, accurate driver hands-off detection is a critical factor for ensuring driving safety. Among the current mainstream hands-off detection technologies, torque-based steering wheel systems suffer from high false alarm rates, capacitivebased steering wheel systems incur high costs, and visual monitoring approaches have blind spots and degraded performance at night. To address these issues, this paper proposes a hands-off detection method based on multi-feature fusion deep learning. The method extracts multi-dimensional features from raw steering wheel torque signals in the time domain, frequency domain, and combined timefrequency domain. A one-dimensional convolutional neural network (CNN) is employed to process the raw signal branch, while a multi-layer perceptron (MLP) is used to process the handcrafted feature branch. The multi-feature fusion network (MFF-Net) then performs feature fusion and classification, directly outputting a binary classification result indicating whether the driver's hands are on or off the steering wheel. Validation on 108 sets of real-world driving data demonstrates that the proposed method achieves a detection accuracy of 99.07%, which meets the precision requirements of L2-level driving assistance systems for hands-off detection. This work provides an optimized solution for low-cost steering wheel hands-off detection and holds considerable engineering application value.

Key words: steering wheel torque; hands-off detection; multi-feature fusion; deep learning