交通事故及其造成的人员伤亡是全球性的社会问题。驾驶员的人为风险因素不仅是造成交通事故的内在关键原因，而且受到外在环境因素的影响。为深入分析内外风险因素相互作用的事故致因机理，本论文在风险平衡理论的框架下设计了一系列离线和在线实验，综合采用行为与生理指标测量驾驶员在具体环境中的“感知-认知-反应”过程，并在此基础上采用模式识别方案预测交通事故的发生。在离线任务方面，本文基于实路视频的风险驾驶行为研究表明：驾驶员间的道路资源竞争导致普遍的风险驾驶行为，而风险驾驶行为既是驾驶员本人的内在个体风险，又是他人的外在环境风险。本文基于离线风险感知的评估结果表明：风险驾驶行为与风险感知水平呈现出负相关关系；而离线风险决策任务和反馈事件相关电位指标的研究结果表明：高风险行为倾向驾驶员对高不确定性和冒险性的选项具有较高的偏好，且其认知决策阶段的脑功能反应表现出了较高的正反馈激励和较低的负反馈抑制。以上离线研究的结果表明，针对驾驶员的感知和认知特点进行干预，可以有效降低其驾驶行为的风险程度，进而降低交通事故机率。而在线任务方面，本文模拟驾驶任务的研究结果表明：驾驶员根据道路环境自发调节其行为，以保持内外风险的平衡状态。本文开展了三类情景下的在线研究：在无风险情景中，驾驶员自由选择其行为；在风险决策情景中，驾驶员作出抢行或让行的决策；在险情情景中，驾驶员进行风险规避。高风险行为倾向驾驶员表现出了较多的抢行决策和区别于其他驾驶员的行为模式；发生事故驾驶员表现出较低的风险预警能力，其导致的生理反应可由皮电上升率指示。以上在线研究的结果，揭示了不同道路环境中内外风险失衡的原因和事故致因机理。在模式识别研究方面，本文采用不同特征组合和不同判别分析模型，验证了采用行为与生理指标实时预测交通事故的可行性。结果表明，包含所选行为与生理指标特征组合的模式识别方案具有较高的准确率和较低的误报率，在此基础上开发的事故预测系统可以综合考虑内外风险因素，准确地预测事故的发生。本论文研究综合采用行为与生理指标测量驾驶过程中的风险因素，揭示了不同环境下驾驶行为背后的深层次心理与生理机理，扩展了现有的风险平衡理论框架。研究提出的事故模式识别方案，可以直接应用于实时在线的交通事故预测，对其他人机系统状态识别研究亦有借鉴作用。
Traffic accidents and related fatalities/injuries are global safety issues. Within the Driver-Vehicle-Environment (DVE) system, traffic accidents are mainly caused by drivers’ risk factors. However, these internal risk factors are also determined by the external factors from the vehicle and road environments, as proposed in the Theory of Risk Homeostasis (TRH). Thus, it becomes imperative to reveal the risk factors from humans in varied road environments and the subsequent accident-caused mechinanism. This dissertation conducted a series of off-line and on-line experimental studies with the advanced approaches of behavioral and physiological measures, to anlysis the underlying mechinanism of traffic accidents from the perspective of drivers’ “perception-cognition-repsonse” process.The off-line studies indicated individual differences between risky and other drivers during perception and cognition process. The naturalistic video-based study showed that the risky driving behaviors (i.e., violations and agrresions) were the major problems on the urban roads of China, which were caused by the competivitve relations between drivers and subsequent interpersonal confilicts. The video-based risk perception test demonstrated the negative correlation between the freuqnency of risky driving and the risk perception level. The decision-making tasks showed that the drivers with higher riskness trait had higher preferences to uncertain and risky choices during decision-making. The feedback-loaced Event Related Potential (ERP) during decision-making further revealed that the neural mechinanism of these risky drivers was less error-revised and more reward-motivated. Thus, it is promsing to improve safety driving behaviors by implementing approprate interventions to these risky factors in drivers’ perception and cogitivtion levels.The on-line studies were conducted in driving simulator contexts. During the driving task, drivers demonstrated varied behaviors to compensate the different risks from road environments. In the low level-risk contexts, drivers decided their dirivng behaviors freely. In the medium-level risk contexts, drivers decided their driving behaviors from alternative choices “go” or “not go”. In the high-level risk contexts, drivers had to adopt emergent responses to avoid crashes. In medium-level risk contexts with decision-making, risky drivers made more “go” decisions than other drivers. In the low-level and medium-level risk contexts, risky drivers demonstrated different behavioral patterns compared with other drivers, as distinguished by longitudinal velocity and horizontal angular velocity. In the high-level risk context, crashed drivers showed low level of anticipatory hazard, which was distinguished by the rising rate of the skin conductance. The above results indicated the key factors from human and accident-caused mechinanism in different road enviorments.In order to predict the accident involvements, this dissertation finally adopted the pattern recognition method with Discriminant Analysis (DA) based on the behavioral and physiological measures. The results showed that the pattern recognition methods with more measures could increase the accuracy and specificity. In this light, it is possible to develop high performance systems to predict the impending accidents and conduct forward interventions accordingly.In summary, this dissertation offers comprehensive guidelines to measure the drivers’ risk factors through the behavioral and physiological measures. The findings provide more precise descriptions in perception and cognition level to explain the accident-caused mechinanism based on the TRH framework. DA and other pattern recognition methods provide practicable approaches to accurately predict impending accidents based on these multiple measures. In addition, it is feasible to extend the behavioral and physiological measures in other areas involving human-machine systems.
