基于自然驾驶数据的分心驾驶行为识别方法

大量证据表明，驾驶人分心是导致交通事故的主要原因之一。当前基于侵入式（如脑电波等）或半侵入式（如视频等）检测驾驶人分心的方法，不仅对驾驶任务造成一定干扰，且受多种环境因素的制约，误报率较高。基于此，只考虑非侵入式车辆运动特征，提出一种基于深度学习的驾驶人分心状态识别方法：首先，从自然驾驶数据集中获得大量的跟驰片段，采用态势感知方法，提取典型的分心驾驶片段，并建立仅包含车辆运动学特征的分心判别指标集；其次，利用梯度提升决策树-递归特征消除算法（GBDT-RFE）和随机森林-递归特征消除算法（RF-RFE）对特征进行重要度排序，得到重要度较高的分心监测指标；最后，采用长短时记忆神经网络（LSTM-NN）实现分心驾驶的分类识别，并与支持向量机和AdaBoost的模型结果进行对比。研究结果表明：LSTM-NN在判别分心或正常状态时F₁分别为89%、91%，高于SVM和AdaBoost对应二分类结果；进行多分类任务时，判别分心情景的平均F₁较SVM和AdaBoost分别提升了12%和7%，不同类别分心识别的误报率在15%以下，说明LSTM-NN能够有效学习分心序列的前后信息，有利于准确估计驾驶人的状态。研究结果可为车辆分心预警系统和驾驶风险倾向性评估提供方法基础。

Detecting Distraction Behavior of Drivers Using Naturalistic Driving Data

A large amount of evidence has shown that driver distraction is one of the main causes of traffic accidents. However, the current methods, mostly based on invasive (e.g., electroencephalogram) or semi-invasive (e.g., video streams) devices to detect driver distraction, cause certain disturbances to driving tasks resulting in higher false-alarm rates. This study presents a deep learning model for distraction recognition based on vehicle kinematic data. Firstly, car-following periods were obtained from the naturalistic driving data. Then, three kinds of distraction periods with only vehicle kinematic features were extracted using scenario-based situation awareness method. The recursive feature elimination method with gradient boosting decision tree and random forest were applied to sort all indicators, remaining 21 indicators of high importance for both. Finally, the long short-term memory neural network (LSTM-NN) was used to achieve the categorical recognition of distracted driving, and compared with support vector machine (SVM) and AdaBoost. The results show that the F₁-scores of LSTM-NN model for two-class classification tasks are 89% and 91%, respectively, which are higher than the other models. In the multi-class classification task, the average macro F₁-scores increased by 12% and 7%, respectively, compared to SVM and AdaBoost, and the false positive rate of all situational distractions are below 15%. These results indicate that LSTM-NN can effectively seize the time-varying context of distracted status. Therefore, this study provides a methodological basis for vehicle distraction warning systems and driving risk propensity assessments.