军事知识服务平台
自动驾驶汽车测试安全员综合能力评价方法
Comprehensive Ability Evaluation Method for Autonomous Vehicle Test Safety Officers
摘要 针对当前自动驾驶测试安全员(以下简称"安全员")综合能力评价缺乏可量化、可操作的评估体系,以及现有驾驶人评价框架无法覆盖"监控-接管-反馈"全职责链条的问题,从自动驾驶测试安全员的岗位特性出发,构建了涵盖反应能力、驾驶能力和描述能力3个维度的综合能力评价指标体系.选取50名一线安全员,设计涵盖无驾驶任务、人工驾驶和自动驾驶3种模式的分阶段实验,采集反应时间、驾驶操作和问题描述等6类13项原始数据.在反应能力评价方面,以反应时间均值、方差和错误次数为核心指标,对高斯混合模型(Gauss-ian mixture model,GMM)与极端随机树(extremely randomized trees,ET)、随机森林(random forest,RF)、极端梯度提升(eXtreme gradient boosting,XGBoost)等6种算法组合进行比较,经"聚类质量-信息保留度-变量利用充分性-分类精度"四级递进筛选,通过轮廓系数(0.82)、赤池信息准则(186.3)、权重分布均匀度和PRF(查准率、查全率、F1 分数)等多维指标综合筛选,确定GMM+ET为反应能力最优评价模型,查准率、查全率和 F1 分数均达到1.00,十折交叉验证平均准确率为0.99.在驾驶能力评价方面,构建以事故发生次数为因变量的多元线性回归模型(调整后 R²=0.552,Prob(F-statistic)<0.01),识别出最大制动值、平均油门程度和最佳速度差值为影响驾驶能力的3个核心指标.在描述能力评价方面,采用多专家分级评分法,从准确度与精炼度2个维度对安全员的问题反馈质量进行量化评定.依据事故归因统计与德尔菲法专家校准,确定反应能力、驾驶能力和描述能力的权重分别为0.402、0.496、0.102.结果表明:3项能力维度相互独立,反应能力与驾驶能力的皮尔逊相关系数为-0.123;驾驶能力得分与综合排名的相关性最强(r=0.87),是影响综合能力的核心因素;GMM+ET融合模型的十折交叉验证平均准确率为0.99、F1 分数为1.00,在变量利用充分性与评价精度上显著优于GMM+RF(F1均值0.63)与k-means+XGBoost(F1均值0.58).
Abstract:
To address the absence of a quantifiable and operational evaluation framework for the comprehensive competence of autonomous vehicle test safety officers,and the inability of existing driver assessment systems to cover the full"monitor-takeover-feedback"responsibility chain,a three-dimensional evaluation index system en-compassing reaction ability,driving ability,and description ability was constructed based on the job characteristics of test safety officers.Fifty frontline test safety officers participated in a staged experiment covering three modes:no-driving task,manual driving,and autonomous driving.Six categories comprising thirteen raw data indicators were collected,including reaction time,driving operations,and problem descriptions.For reaction ability evalua-tion,reaction time mean,variance,and error count were adopted as core indicators.Six algorithmic combinations in-volving Gaussian Mixture Model(GMM),Extremely Randomized Trees(ET),Random Forest(RF),and eXtreme Gradient Boosting(XGBoost)were preliminarily compared and screened through a four-level progressive strategy of clustering quality,information retention,variable utilization sufficiency,and classification accuracy,based on the silhouette coefficient(0.82),Akaike information criterion(AIC)of 186.3,weight distribution uniformity,and PRF(Precision,Recall,F1-score).The GMM+ET fusion model was identified as the optimal evaluation model,achiev-ing precision,recall,and F1-score all equal to 1.00,with a mean accuracy of 0.99 under ten-fold cross-validation.For driving ability evaluation,a multiple linear regression model(adjusted R²=0.552,Prob(F-statistic)<0.01)with accident frequency as the dependent variable was constructed,identifying maximum braking value,average throttle input,and optimal speed deviation as the three core influencing indicators.For description ability evaluation,a multi-expert grading method was employed to quantify the accuracy and conciseness of safety officers'problem feedback.Based on accident attribution statistics and Delphi method expert calibration,the weights of reaction abili-ty,driving ability,and description ability were determined as 0.402,0.496,and 0.102,respectively.The results dem-onstrate that the three ability dimensions are mutually independent,with a Pearson correlation coefficient of-0.123 between reaction and driving abilities.Driving ability score exhibits the strongest correlation with comprehensive ranking(r=0.87),representing the core factor influencing overall competence.With a ten-fold cross-validation mean accuracy of 0.99 and F1-score of 1.00,the GMM+ET fusion model significantly outperforms GMM+RF(mean F1-score of 0.63)and k-means+XGBoost(mean F1-score of 0.58)in variable utilization adequacy and evaluation ac-curacy.
刊 名 交通信息与安全 ISTICPKU
年,卷(期): 2026, 44(4)
分类号: U491.2
机标分类号: U471.15 D923 TP391.41
在线出版日期: 2026年11月6日
基金项目: 21YJCZH060:教育部人文社会科学研究项目KM202410038001:北京市教委科技一般项目,