基于强弱感知设计的驾驶员危险感知状态识别模型研究

Research on the Driver's Hazard Perception State Recognition Model Based on Strength and Weakness Perception Design

驾驶员危险感知能力对于预防和减少道路交通事故具有重要作用。针对目前研究中危险感知的特征向量表征不统一、算法对实际问题可解释性不足的弊端,通过人为控制的方式,设计了基于危险源、显隐性、强弱危险感知状态3个维度的3×2×2实验方案,实现对危险感知的预定分级;设计配对T检验和Wilcoxon符号秩检验结合的方式实现强弱感知状态下特征差异性的量化比较;建立基于十倍交叉参数调优SVM算法的危险感知状态二分类模型。研究结果表明:驾驶员在强危险感知状态下对危险的反应更主动,倾向于避免危险,而非紧急避险,并保持更低的车速,更倾向油门控制而非制动控制,注视和眼跳行为增加;驾驶员在隐性危险源场景中的操纵力度和频次更强,危险感知与显隐性影响程度及危险源类型有关,摩托车类危险源差异最大,行人类危险源差异最小;在C=1,γ=0.1,选择车头时距、车速标准差、最大制动踏板力、加速度标准差、预减速时间、车速均值、油门开度标准差、跳视次数、注视次数作为特征时,SVM模型具有最好的性能,准确率为89.2%,精准率为90.6%,召回率为87.8%,F1值为0.888。XGBoost模型对弱感知的识别能力低于SVM模型。该研究对于驾驶员危险感知状态的量化评估具有显著的指导意义。

Driver hazard perception plays an important role in preventing and reducing road traffic accidents.For the disadvantages of inconsistent representation of feature vectors of hazard perception and insufficient interpretability of algorithms to practical problems,in this paper 3×2×2 experimental scenarios of three dimensions in terms of danger resource,overt and covert hazard scenes,strong and weak hazard perception state through artificial control to realize the predetermined classification of hazard perception.A combination of paired T-test and Wilcoxon signed-rank test is designed to quantitatively compare the difference of features in the state of strong and weak perception.A binary classification of hazard perception state based on 10-fold cross-parameter tuning SVM algorithm is proposed.The results show that drivers are more active in reacting to danger in the state of strong hazard perception,tending to avoid danger rather than emergency avoidance,while maintaining a lower speed,preferring throttle control rather than brake control,with increase of gaze and saccade behaviors.In the scene of covert hazard source,the driver's manipulation is stronger and more frequent,and the level of HP affected by the overt and covert hazard is related to the type of hazard,with highest level by motorcycle and lowest by human.At C=1,γ=0.1,the SVM model has the best performance with the accuracy of 89.2%,the precision of 90.6%,the recall of 87.8%,and the F1 value of 0.888 when the time headway,standard deviation of vehicle speed,maximum brake pedal force,standard deviation of acceleration,pre-deceleration time,mean vehicle speed,standard deviation of throttle opening,number of saccades,number of fixations are selected as features.The model of XGBoost has lower recognition ability for weak perception than the model of SVM.This study has significant guiding significance for the quantitative evaluation of drivers'hazard perception state.