准确识别车辆当前驾驶行为模式是自动驾驶领域亟待解决的技术问题。为实现驾驶行为模式精准解析,提高模型识别精度和可靠性,通过开展自然驾驶试验,采集高速工况下20名驾驶人的驾驶行为数据及视觉特性数据等多源参数信息,分析4类典型驾...准确识别车辆当前驾驶行为模式是自动驾驶领域亟待解决的技术问题。为实现驾驶行为模式精准解析,提高模型识别精度和可靠性,通过开展自然驾驶试验,采集高速工况下20名驾驶人的驾驶行为数据及视觉特性数据等多源参数信息,分析4类典型驾驶行为模式(自由行驶、跟车、左换道、右换道)运行规律及多源参数耦合特性。基于主成分分析法确定4类驾驶行为模式表征指标集,使用支持向量机、随机森林决策树算法建立驾驶行为模式识别模型,通过学习训练,分析比较模型识别结果,对识别效果较好的模型进一步优化,分析优化模型对4类驾驶行为模式识别精度的时序性变化特征。研究结果表明:支持向量机模型、随机森林决策树模型、基于多层感知器神经网络的随机森林优化模型总体识别精度分别为89.4%、90.5%、91.9%;4类驾驶行为模式的AUC(area under the curve)值均高于0.93,可较好地识别车辆当前驾驶行为模式。此外,随机森林优化模型对4类驾驶行为模式的识别精度均随时间推移,呈现先增长后趋于稳定的变化态势,且同一时刻的自由行驶及跟车模式识别精度高于向左及向右换道模式。研究结果可为高级别自动驾驶系统决策及控制策略的制定提供理论基础和技术支持。展开更多
It is difficult to model human behavior because of the variability in driving styles and driving skills. However, for some driver assistance systems, it is necessary to have knowledge of that behavior to discriminate ...It is difficult to model human behavior because of the variability in driving styles and driving skills. However, for some driver assistance systems, it is necessary to have knowledge of that behavior to discriminate potentially hazardous situations, such as distraction, fatigue or drowsiness. Many of the systems that look for driver distraction or drowsiness are based on intrusive means （analysis of the electroencephalogram--EEG） or highly sensitive to operating conditions and expensive equipment （eye movements analysis through artificial vision）. A solution that seeks to avoid the above drawbacks is the use of driving parameters This article presents the conclusions obtained after a set of driving simulator tests with professional drivers with two main objectives using driving variables such as speed profile, steering wheel angle, transversal position on the lane, safety distance, etc., that are available in a non-intrusive way： （1） To analyze the differences between the driving patterns of individual drivers; and （2） To analyze the effect of distraction and drowsiness on these parameters. Different scenarios have been designed, including sequences with distractions and situations that cause fatigue. The analysis of the results is carried out in time and frequency domains in order to identify situations of loss of attention and to study whether the evolution of the analyzed variables along the time could be considered independent of the driver.展开更多
文摘准确识别车辆当前驾驶行为模式是自动驾驶领域亟待解决的技术问题。为实现驾驶行为模式精准解析,提高模型识别精度和可靠性,通过开展自然驾驶试验,采集高速工况下20名驾驶人的驾驶行为数据及视觉特性数据等多源参数信息,分析4类典型驾驶行为模式(自由行驶、跟车、左换道、右换道)运行规律及多源参数耦合特性。基于主成分分析法确定4类驾驶行为模式表征指标集,使用支持向量机、随机森林决策树算法建立驾驶行为模式识别模型,通过学习训练,分析比较模型识别结果,对识别效果较好的模型进一步优化,分析优化模型对4类驾驶行为模式识别精度的时序性变化特征。研究结果表明:支持向量机模型、随机森林决策树模型、基于多层感知器神经网络的随机森林优化模型总体识别精度分别为89.4%、90.5%、91.9%;4类驾驶行为模式的AUC(area under the curve)值均高于0.93,可较好地识别车辆当前驾驶行为模式。此外,随机森林优化模型对4类驾驶行为模式的识别精度均随时间推移,呈现先增长后趋于稳定的变化态势,且同一时刻的自由行驶及跟车模式识别精度高于向左及向右换道模式。研究结果可为高级别自动驾驶系统决策及控制策略的制定提供理论基础和技术支持。
文摘It is difficult to model human behavior because of the variability in driving styles and driving skills. However, for some driver assistance systems, it is necessary to have knowledge of that behavior to discriminate potentially hazardous situations, such as distraction, fatigue or drowsiness. Many of the systems that look for driver distraction or drowsiness are based on intrusive means （analysis of the electroencephalogram--EEG） or highly sensitive to operating conditions and expensive equipment （eye movements analysis through artificial vision）. A solution that seeks to avoid the above drawbacks is the use of driving parameters This article presents the conclusions obtained after a set of driving simulator tests with professional drivers with two main objectives using driving variables such as speed profile, steering wheel angle, transversal position on the lane, safety distance, etc., that are available in a non-intrusive way： （1） To analyze the differences between the driving patterns of individual drivers; and （2） To analyze the effect of distraction and drowsiness on these parameters. Different scenarios have been designed, including sequences with distractions and situations that cause fatigue. The analysis of the results is carried out in time and frequency domains in order to identify situations of loss of attention and to study whether the evolution of the analyzed variables along the time could be considered independent of the driver.
