Fu Chuanyun, Lu Zhaoyou, Liu Huahua, Wumaierjiang Ayinigeer
School of Transportation Science and Engineering, Harbin Institute of Technology, China; Department of Civil Engineering, The University of British Columbia, Canada.
School of Transportation Science and Engineering, Harbin Institute of Technology, China.
Accid Anal Prev. 2025 Jul;217:108065. doi: 10.1016/j.aap.2025.108065. Epub 2025 Apr 28.
Dynamic short-term crash risk prediction derived from traffic conflicts can provide significant support for proactive safety management at signalized intersections. Especially after the formation of emerging mixed traffic flow, an accurate prediction of future crash risk can help conceive proactive crash prevention measures. However, the precision of crash risk estimation at signalized intersections with emerging mixed traffic flow is still subject to doubt, largely attributable to the lack of an exclusive conflict indicator. This situation presents considerable challenges to the dynamic short-term crash risk prediction at signalized intersections with emerging mixed traffic flow. Therefore, this study performs dynamic short-term crash risk prediction from traffic conflicts at signalized intersections with emerging mixed traffic flow by combining the non-stationary generalized extreme value (GEV) model and the self-attention mechanism-based online learning long short-term memory (SAM-OL-LSTM) approach. A novel conflict indicator, the time to avoid a crash (TTAC), is developed to describe traffic conflicts in the emerging mixed traffic flow. Based on TTAC, a non-stationary GEV model that considers acceleration variance as a covariate is developed to calculate the value at risk (VaR) for each minute, which is used to dynamically quantify crash risk at signalized intersections. Afterwards, the SAM-OL-LSTM approach that considers traffic volume and the uncertainty in vehicle speed distribution as two input features is proposed to dynamically predict the VaR for the future one minute based on the VaR time series data of the prior five minutes. The results indicate that: i) the proposed SAM-OL-LSTM approach outperforms baseline approaches under various MPRs in terms of prediction accuracy; ii) the application of VaR facilitates a dynamic quantification of the crash risk at an intra-minute temporal resolution; iii) the developed TTAC exhibits a strong capability in identifying traffic conflicts in the emerging mixed traffic flow at signalized intersections. The findings of this study can provide a theoretical foundation for proactive traffic control considering the future crash risk in the emerging mixed traffic flow at signalized intersections.
基于交通冲突的动态短期碰撞风险预测可为信号交叉口的主动安全管理提供重要支持。特别是在新兴混合交通流形成后，准确预测未来碰撞风险有助于构思主动预防碰撞的措施。然而，新兴混合交通流下信号交叉口碰撞风险估计的精度仍受到质疑，这在很大程度上归因于缺乏专门的冲突指标。这种情况给新兴混合交通流下信号交叉口的动态短期碰撞风险预测带来了相当大的挑战。因此，本研究通过结合非平稳广义极值（GEV）模型和基于自注意力机制的在线学习长短期记忆（SAM-OL-LSTM）方法，对新兴混合交通流下信号交叉口的交通冲突进行动态短期碰撞风险预测。开发了一种新颖的冲突指标——避免碰撞时间（TTAC），以描述新兴混合交通流中的交通冲突。基于TTAC，开发了一种将加速度方差作为协变量的非平稳GEV模型，用于计算每分钟的风险价值（VaR），该值用于动态量化信号交叉口的碰撞风险。随后，提出了将交通流量和车速分布的不确定性作为两个输入特征的SAM-OL-LSTM方法，基于前五分钟的VaR时间序列数据动态预测未来一分钟的VaR。结果表明：i）所提出的SAM-OL-LSTM方法在预测准确性方面在各种MPR下均优于基线方法；ii）VaR的应用有助于在分钟内时间分辨率下对碰撞风险进行动态量化；iii）所开发的TTAC在识别新兴混合交通流下信号交叉口的交通冲突方面表现出很强的能力。本研究结果可为考虑新兴混合交通流下信号交叉口未来碰撞风险的主动交通控制提供理论基础。
