基于XGBoost算法的危险场景驾驶行为模式分析及安全评估

危险感知能力对驾驶人的驾驶行为模式具有重要影响。为准确评估驾驶人的危险感知能力、提升危险感知水平判别的准确度，提出了基于模拟驾驶技术的危险感知能力影响分析方法和基于极端梯度提升树（XGBoost）算法的危险感知水平判别模型。通过设计3种常见交通冲突场景，采集模拟驾驶中驾驶人的多维度驾驶行为特征数据，并分析危险感知能力与驾驶行为的相关关系。通过模拟实验发现:驾驶人对行人的危险感知能力较弱，易发生碰撞事故；驾驶人在危险场景中的车速（p＝0.01）、制动反应位置（p < 0.01）以及反应时间（p < 0.01）与危险感知水平之间存在显著负相关关系。在相关性分析的基础上，利用XGBoost算法识别能反映驾驶人危险感知能力的重要特征变量，并构建以制动反应位置、反应时间、车速、刹车深度，以及加速度为指标的驾驶人危险感知水平判别模型；通过与LightGBM、支持向量机（SVM），以及逻辑回归（LR）等算法分类预测性能的对比分析，评价危险感知模型的判别精度，结果表明:基于XGBoost算法的危险感知水平判别模型的判别准确率为84.8%、F1值为83.4%、AUC值为0.959，优于LightGBM（准确率为78.8%、F1值为76.7%、AUC值为0.924）、SVM（准确率为57.6%、F1值为42.2%、AUC值为0.859），以及LR算法（准确率为69.7%、F1值为65.5%、AUC值为0.836）。所提方法可为判别驾驶人危险感知能力及其对驾驶行为模式的影响提供可靠手段。 

An Analysis of Driving Behavior Model and Safety Assessment Under Risky Scenarios Based on an XGBoost Algorithm

Hazard perception is a critical factor of a driving behavior model. A simulator-based method and an extreme gradient boosting tree(XGBoost)algorithm are proposed, in order to study the impacts of hazard perception on driving behaviors and improve the accuracy of hazard perception. Three typical scenarios of traffic conflicts are simulated, and a large amount of driving behavior data are collected. The correlation between hazard perception and driving behavior models is discussed under the three scenarios. The correlation analysis reveals that when hazard perception(e.g., dangerous behaviors of pedestrians)is weak, and the vehicle speed(p=0.01), braking reaction position(p < 0.01), and reaction time(p < 0.01)are significantly negatively correlated with the drivers'hazard perception. Based on the correlation analysis, the XGBoost algorithm is used to identify important features determining the capability of hazard perception of drivers. Then, a discriminant model of hazard perception is proposed with following the indicators, such as braking reaction position, reaction time, vehicle speed, braking depth, and acceleration. Compared the proposed method with Light Gradient Boosting Machine(LightGBM), Support Vector Machine(SVM), and Logistic Regression(LR)algorithms, it is found that the accuracy of the XGBoost-based method is 84.8%, its F1-score is 83.4%, and the area under the receiver operating characteristic Curve(AUC)is 0.959, which is better than the LightGBM(accuracy is 78.8%, F1-score is 76.7%, and AUC is 0.924), SVM(accuracy is 57.6%, F1-score is 42.2%, and AUC is 0.859)and LR algorithm(accuracy is 69.7%, F1-score is 65.5%, and AUC is 0.836). In conclusion, the proposed method can provide a more reliable way for understanding the capability of hazard perception of drivers and its impacts on driving behavior models.