Application of a Rule-Based Approach in Real-Time Crash Risk Prediction Model Development Using Loop Detector Data

<b>Objectives</b>: There is a growing trend in development and application of real-time crash risk prediction models within dynamic safety management systems. These real-time crash risk prediction models are constructed by associating crash data with the real-time traffic surveillance data (e.g., collected by loop detectors). The main objective of this article is to develop a real-time risk model that will potentially be utilized within traffic management systems. This model aims to predict the likelihood of crash occurrence on motorways.<b>Methods</b>: In this study, the potential prediction variables are confined to traffic-related characteristics. Given that the dependent variable (i.e., traffic safety condition) is dichotomous (i.e., “no-crash” or “crash”), a rule-based approach is considered for model development. The performance of rule-based classifiers is further compared with the more conventional techniques like binary logistic regression and decision trees. The crash and traffic data used in this study were collected between June 2009 and December 2011 on a part of the E313 motorway in Belgium between Geel-East and Antwerp-East exits, on the direction toward Antwerp.<b>Results</b>: The results of analysis show that several traffic flow characteristics such as traffic volume, average speed, and standard deviation of speed at the upstream loop detector station and the difference in average speed on upstream and downstream loop detector stations significantly contribute to the crash occurrence prediction. The final chosen classifier is able to predict 70% of crash occasions accurately, and it correctly predicts 90% of no-crash instances, indicating a 10% false alarm rate.<b>Conclusions</b>: The findings of this study can be used to predict the likelihood of crash occurrence on motorways within dynamic safety management systems.

<b>研究目标</b>: 动态安全管理系统（Dynamic Safety Management Systems）领域内，实时碰撞风险预测模型（Real-time Crash Risk Prediction Models）的开发与应用正呈现日益增长的趋势。此类模型通过将碰撞事故数据与实时交通监测数据（例如由环形检测器（Loop Detectors）采集的数据）相关联进行构建。本文的核心研究目标为开发一款可应用于交通管理系统的实时风险模型，该模型旨在预测高速公路（Motorways）上碰撞事故发生的可能性。
<b>研究方法</b>: 本研究中，潜在预测变量仅限定于交通相关特征参数。由于因变量（即交通安全状况）为二分类变量（仅包含"无碰撞"与"有碰撞"两类），因此本研究采用基于规则的方法开展模型开发。研究进一步将基于规则分类器的模型性能与二元逻辑回归、决策树等传统技术进行对比。本研究使用的碰撞事故与交通数据采集于2009年6月至2011年12月期间，采集地点为比利时E313高速公路吉勒东（Geel-East）与安特卫普东（Antwerp-East）收费站之间朝向安特卫普方向的路段。
<b>研究结果</b>: 分析结果表明，多项交通流特征参数——如上游环形检测器站点的交通流量、平均车速及车速标准差，以及上下游环形检测器站点的平均车速差值——对碰撞事故发生预测具有显著贡献。最终选定的分类器可准确预测70%的碰撞事故场景，同时对90%的无碰撞场景预测正确，假警报率为10%。
<b>研究结论</b>: 本研究的发现可用于在动态安全管理系统中预测高速公路上碰撞事故发生的可能性。