Abstract: To address the issue that single-modal systems in driver monitoring systems cannot simultaneously achieve both early recognition and stability, this paper proposes a driving behavior recognition method that considers the temporal relationship between gaze and steering wheel grip. Based on the behavioral rule that "perception precedes manipulation" during driving, gaze and grip signals are synchronously acquired and time-aligned. On this basis, a sliding window is adopted to perform temporal reconstruction of multi-source signals, extract feature parameters reflecting visual attention allocation and manipulation readiness, and construct a random forest model for driving behavior classification. Meanwhile, cross-correlation analysis is employed to quantify the time-lag relationship between gaze and grip, and this temporal characteristic is incorporated into the feature fusion process. Cross-subject validation is conducted based on simulated driving data from 50 drivers. The experimental results show that under lane-changing and turning conditions, gaze features stably lead grip features by approximately 1.0～1.5 s. The fusion model achieves an area under the curve (AUC) of 0.94, outperforming single-modal models in recognition performance. The multi-modal method incorporating temporal coupling relationships can effectively improve the accuracy and robustness of driving behavior recognition, providing a basis for active safety warning.

Key words: driving behavior recognition; multi-modal fusion; temporal coupling; random forest; driver monitoring system

摘要： 针对驾驶员监控系统中单一模态难以同时兼顾识别提前性与稳定性的问题，文章提出 一种考虑视线与方向盘握力时序关系的驾驶行为识别方法。基于驾驶过程中“感知先于操 控”的行为规律，对视线与握力信号进行同步采集与时间对齐。在此基础上，采用滑动窗口 对多源信号进行时序重构，提取反映视觉注意分配与操控准备状态的特征参数，并构建随机 森林模型实现驾驶行为分类。同时，通过互相关分析量化视线与握力之间的时滞关系，将该 时序特性引入特征融合过程。基于 50 名驾驶员的模拟驾驶数据开展跨被试验证，试验数据表 明，在换道与转弯工况下，视线特征相对握力特征稳定领先约 1.0～1.5 s；融合模型曲线下面 积（AUC）达到 0.94，识别性能优于单一模态模型。引入时序耦合关系的多模态方法能够有 效提升驾驶行为识别的准确性与鲁棒性，为主动安全预警提供依据。

关键词: 驾驶行为识别；多模态融合；时序耦合；随机森林；驾驶员监控系统