驾驶分心综述

Review on driving distraction

构建了文献质量评价指标体系,筛选出288篇相关文献,综合分析了文献数据获取方式、指标选取、识别方法和研究结论;以驾驶行为为主要研究对象,结合统计学方法,系统阐述了驾驶分心试验数据的获取方式,总结了其多样性和两极化的原因;梳理了驾驶分心识别指标的研究成果,总结了其使用效能和优缺点;对比了不同驾驶分心识别模型的准确率,分析了其差异性的根源;提出了未来驾驶分心数据获取方式、指标选取、识别方法的研究趋势。分析结果表明:试验是驾驶分心数据获取的主要方式,自然驾驶数据集、视频录像是新的数据获取方式,路边观察和调查问卷的数据获取方式关注度较低;对照、跟车、超车、变道场景及添加其他危险事件的复杂度较高的场景是研究较多的驾驶分心场景;驾驶分心次任务的设置说明目前驾驶分心研究的类型和主题较集中;融合指标是使用频次最高的驾驶分心识别指标,且以驾驶绩效指标和眼动指标、驾驶绩效指标和反应指标这两类融合指标较多,驾驶绩效指标是使用最多的单类指标;支持向量机模型是使用最多的驾驶分心识别模型,但识别准确率标准差较大,性能不稳定,深度学习算法模型的识别准确率较高,且稳定性好;未来的驾驶分心研究需均衡研究主题,扩展人机共驾的分心场景,深化驾驶分心类型研究,构建标准化指标体系和选取原则,强化模型构建识别不同类型和严重程度的驾驶分心。

An indicator system for evaluating the quality of literatures was established. Based on this system and considering driving behavior as the main focus of this research, 288 relevant papers were selected, and their data acquisition methods, indicator selections, detection methods, and research conclusions were comprehensively analyzed. Taking driving behavior as the main research object, a method of obtaining test data on driving distractions was systematically derived combined with statistical methods, and the reasons for the diversity and polarization in the obtained data were summarized. The research results of different driving distraction indicators were categorized, and the efficiency, advantages, and disadvantages of these indicators were summarized. The accuracies of different driving distraction detection models were compared, and the root causes of their differences were analyzed. Future research trends of driving distraction data acquisition methods, indicator selections, and detection methods were proposed. Analysis results show that experimental tests are the primary methods for obtaining driving distraction data. Natural driving datasets and video recordings have been proposed as new methods of data acquisition, data acquisition methods of roadside observations and surveys have received less attention. Comparison scenario, vehicle following scenario, overtaking scenario, lane changing scenario, and relatively more complex scenarios involving other dangerous events are the most extensively studied driving distraction scenarios. The setting of driving distraction sub-tasks indicates that current research on driving distraction has focused on several types and topics. Fusion indicators, generally including driving performance and eye movement indicator, and driving performance and reaction indicator, are the most frequently used in driving distraction. Driving performance is the most commonly used single indicator. Support vector machine model is the most commonly used driving distraction detection model, while the standard deviation of detection accuracy is large, and this model is also unstable. In contrast, the detection accuracy of a deep learning algorithm-based model is high, and its stability is good. Future research on driving distraction should balance research topics, expand distraction scenarios to human-machine co-driving, further investigate the types of driving distractions, construct a standardized indicator system and selection principles, and strengthen model construction to detect different types and determine the severity of driving distractions. 11 tabs, 1 fig, 96 refs.