高速公路路侧雷达布设间距对交通事故风险评估精度的影响

高速公路通过布设毫米波雷达等新型检测设备，实现交通状态的精准感知，并为主动交通管控提供支撑。然而检测设备布设成本高，其布设间距需综合考虑成本约束和交通状态感知成效。为探究路侧毫米波雷达布设间距对交通事故风险评估精度的影响，基于浙江省沪杭甬高速公路的实证数据开展研究。构建事故风险评估深度森林模型(deep forest，DF)，应用滑动时空窗提取交通运行特征，并通过多层级联随机森林的集成建立交通运行特征与事故风险的关联关系；考虑路侧毫米波雷达感知范围，构建不同雷达布设间距下的交通运行数据集，开展布设间距对事故风险评估模型精度的敏感性分析。研究结果表明：DF模型曲线面积值(area under curve，AUC)为0.849，事故样本分类准确率为80.9%，高于传统的卷积神经网络模型(AUC值为0.741，准确率为75.2%)、随机森林模型(AUC值为0.715，准确率为70.8%)；雷达布设间距与事故风险评估精度呈反比关系，且密集布设下模型精度提升的边际效应递减，当布设间距由1 500 m缩减至750 m时，事故风险评估模型AUC值呈显著上升趋势，由0.794提升至0.853，布设间距由750 m缩减至250 m时，AUC值无明显变化。综上，雷达布设间距为750 m可平衡布设成本和事故风险评估精度，成果可为高速公路车道级交通状态感知系统的规划设计提供决策依据。

Effects of Spacing of Highway Roadside Millimeter-wave Radar Detectors on the Accuracy of a Crash Risk Evaluation Model

Freeways equipped with new sensing equipment such as millimeter-wave radar detectors can accurately monitor traffic operation and well support active traffic management measures. However, due to the high deployment expenditure, there is a need to consider the cost constraints and the effectiveness of traffic state detection. To investigate the impacts of millimeter-wave radar deployment spacing on crash risk evaluation performance, this study is conducted based on the empirical data of the Hangshaoyong highway in Zhejiang Province. A crash risk evaluation model based on deep forest (DF) is developed. Specifically, sliding spatio-temporal windows are employed to extract the features of traffic operation while the correlation relationships between the features and crash risk are established through the integrations of multi-layer cascaded random forests. Considering the sensing range of the millimeter-wave radar detectors, multiple traffic operation datasets are developed by assuming different deployment spacings. Sensitivity analyses of radar deployment spacing on the evaluation accuracy of crash risk are conducted. Analyses results show that: The area under curve (AUC) of DF model is 0.849 with 80.9% recall on crash samples, which is higher than traditional convolutional neural network model (AUC is 0.741, recall is 75.2%) and random forest model (AUC is 0.715, recall is 70.8%). An inverse relationship between radar deployment spacing and evaluation accuracy of crash risk is captured, and the marginal effects of the improvement to the model accuracy decreases under dense deployment conditions. If the radar deployment spacing is reduced from 1 500 m to 750 m, the AUC of crash risk evaluation model shows a substantial increase (from 0.794 to 0.853), but there is no obvious change in AUC values when the radar deployment spacing is reduced from 750 m to 250 m. In conclusion, the radar deployment spacing of 750 m can balance the deployment cost and the evaluation performance of crash risk, which could be used to support the decisions related to the installment of traffic sensing equipment.