Abstract: The recognition of driver distraction states is the basis of driver distraction warning. Because visual distraction poses a greater threat to driving safety than cognitive distraction, this paper studies the recognition of both types of distraction. A simulated driving test was designed for normal driving and different types of distraction. The 1-back task was used to create the cognitive distraction and the phone tasks was used to create the visual distraction of drivers. Then the driving performance, eye movement, and head movement were collected and extracted. The sequential backward selection was used for feature selection, and the grid search method was applied to determine the best time window and model parameters for driver distraction recognition. The results show that the established model based on random forest achieves 94.07% macro accuracy, 93.89% macro recall, and 93.98% macro F1 value on the test set. The classification performance is better than the two traditional methods for comparison, indicating that the model can accurately classify the three states of drivers. From the ranking results of the feature importance output by the random forest model and the classification results of the model with different types of features as input, it is found that driver's eye movement and head movement features are more important for the recognition of different distractions of drivers. This study provides a basis for the driver distraction warning system to determine the risk level according to the type of distraction.

Key words: traffic engineering, type of distraction, recognition model, driving performance, eye movement and head movement

摘要： 驾驶人分心状态判别是实现分心预警的基础。由于视觉分心相比认知分心对行车安全具有更大威胁，本文针对驾驶人不同分心类型的识别展开研究，设计了驾驶人两种分心类型和正常驾驶下的模拟驾驶试验，利用1-back任务和看手机任务分别诱导驾驶人产生认知分心和视觉分 心，采集并提取驾驶绩效、眼动及头动特征，采用序列后向选择算法进行特征优选，运用网格搜索确定分心识别的最佳时间窗口及模型参数。结果表明，基于随机森林所构建模型在测试集上取 得了94.07%的宏精准率、93.89%的宏召回率和93.98%的宏F1值，分类表现优于两种比较方法，说明模型能够准确地对驾驶人的3种状态进行分类。根据随机森林模型的特征重要性排序结果以 及采用不同类型特征作为输入训练模型的分类结果发现，驾驶人眼动及头动特征对驾驶人分心类型的识别更为重要。本文研究可为分心预警系统根据分心类型判定风险等级提供基础。

关键词: 交通工程, 分心类型, 识别模型, 驾驶绩效, 眼动及头动