Considering human-vehicle handover to predict autonomous vehicle conflict risk: A deep learning method for radar-video integrated data
Abstract
-
使用交通流特征的实时冲突预测模型（之间的关系）
-
使用HighD 数据集（经验数据），使用虚拟检测器方法进行 1. 交通特征提取 2. 两步框架 -> 轨迹数据分析的新方法。
- 对具有均值和方差异质性的随机参数 logit 模型的探索性研究
- 几种机器学习方法（eXtreme Gradient Boosting (Boosting)、Random Forest (Bagging)、Support Vector Machine (Single-classifier)和多层感知器）的比较研究
-
结论：
- 交通流特征对冲突发生概率有显着影响。
- 考虑**[平均异质性](# 异质性 Heterogeneity )**的统计模型优于对应变量，车道差异变量对【车道变量+车道差异变量】的随机参数均值有显着影响。
- 在欠采样数据集上训练的 eXtreme Gradient Boosting 是最佳模型：AUC 0.871，精度 0.867
- 模型对冲突阈值敏感 -> 冲突风险预测应同时考虑主体车道特征和车道差异特征
Introduction
conflict-prone modols are preferred
-
previews studies: based on crash data, aiming at estimating crash risk -
real-time traffic flow characteristics - ( effective ) -> in real-time crash risk prediction -
limitations in real-time crash risk prediction:crash data error ( inaccurate time, long collection time, temporal or spatial instability of accident data -> limited transferability)not enough historical crashes alongwith traffic flow data
-
more general objective: crash-prone and conflict-prone -
Conflict modelingis are more proactive: don't need crash data and requires shorter data collecting time
real-time conflict-based study been neglected and need to be paid attention
- use microscopic kinematics reaction instead of traffic flow features
- known little about the conflict mechanism and inherent heterogeneity from a traffic flow perspective
aim:
-
conflicts - ( bridge ) - dynamic traffic flow characteristics
-
traditional crash risk prediction - ( extend ) -> conflict-based
objectives: 从宏观交通特征的角度理解冲突机制，以及开发具有高预测性能 和 现实就业适用性的冲突风险模型
贡献：
- 提出了一种利用虚拟检测器方法提取交通 流数据的新方法，并提出了一种用于轨迹数据分析的两步框架;
- 通过探索它们之间的关联，将交通流 和微观冲突联系起来，同时考虑未观察到的异质性;
- 总结了机器学习方法、不平衡学习技术以及冲突阈 值和交通特征选择的敏感性分析的经验教训，以供未来研究。
Literature review
Real-time crash risk prediction
- resl-time risk studies is very diffrent from traditional crash prediction analysis
- real-timecrash risk study is a [disaggregate](# disaggregate and aggregate data) safety study, which considers each crash as an observation and focuses on conditions where a crash is more likely to occur. (安全关键事件附近的空间和时间上的动态交通特征)
- Traditional crash frequency analysis, typical aggregate study, which aims at identifying a specific region where more crashes are likely to occur( 通过泊松、负二项式、泊松-正态等变量数据模型来估计碰撞频率 )
Common analytic framework in real-time crash analysis
five critical steps:
- 基于现有知识提出交通参数的描述性统计，定义交通特征聚合（traffic feature aggregation）的时间窗口
- 基于历史碰撞数据和交通数据，构建包含碰撞观测和正常观测的初步数据集
- 按照指定的碰撞案例和正常案例的比例对初步数据集进行重新采样
- 考虑二元分类问题并开发模型（分类器）以识别具有高碰撞可能性的交通状况
- 结合实际应用，通过多指标评估模型性能
two methods: machine learning models are better than statistical models
- need of the consideration of heterogeneity
- ML are superior in capturing the high-dimensional and non-linear relationship between input features and outcomes
- ML - ( black-box issue ) -> ,weak inferential ability and cannot obtain a thorough understanding of the mechanism of crashes or conflicts
两步走的框架工作:
- 统计模型来确认交通流量和冲突发生的关联，并了解相应的冲突机制
- 机器学习方法来实时预测冲突风险
Real-time conflict-based studies
terminology
-
异质性 Heterogeneity
- 一个变量X对另一个变量Y的影响可能因个体而异: 多上一年学让张三的收入增加了1000元，让李四的收入增加了1200元，那么教育年限对收入的影响就存在异质性；
-
异方差：Heteroskedasticity
- 在变量X的不同水平上，变量Y取值的波动大小可能不同。例：所有小学毕业的人，有的做了老板年入百万，有的成为工薪阶层年入几万——在六年教育水平上，收入取值的波动很大。所有大学毕业的人，都能找到不错的工作，收入多的年赚百万，收入低的也有几十万——在十六年的教育水平上，收入取值的波动较小。
-
disaggregate and aggregate data
