Department of Civil and Environmental Engineering, Louisiana State University, Patrick Taylor Hall, Baton Rouge, LA, 70803, USA.
Accid Anal Prev. 2021 May;154:106090. doi: 10.1016/j.aap.2021.106090. Epub 2021 Mar 16.
Highway work zones are most vulnerable roadway segments for congestion and traffic collisions. Hence, providing accurate and timely prediction of the severity of traffic collisions at work zones is vital to reduce the response time for emergency units (e.g., medical aid), accordingly improve traffic safety and reduce congestion. In predicting the severity of traffic collisions, previous studies used different statistical and machine learning models with accuracy as the main evaluating factor. However, the performance of these models was generally not good, especially on fatal and injury crashes. Also, looking into the prediction accuracy only is misleading. This paper aims to propose a novel deep learning-based approach with a customized f1-loss function to predict the severity of traffic crashes. Underlying this objective is to compare the results of deep learning models with machine learning model considering two performance indicators, namely precision, and recall. The data used in the analysis include a sample of traffic crashes that occurred at work zones in Louisiana from 2014 to 2018. This dataset includes valuable information (features) related to road, vehicle, and human factors affecting the occurrence and severity of those crashes. The proposed methodology is based on transforming these features/variables into images. Image transformation is conducted using a nonlinear dimensionality reduction technique t-SNE and convex hull algorithm. A CNN based deep learning algorithm with a customized loss function was used to directly optimize the model for precision and recall. The results showed improved performance in predicting the crash severity of fatal and injury crashes using the deep learning approach, which can help to improve traffic safety as well as traffic congestion at work zones and possibly other roadways segments.
高速公路工作区是最容易发生拥堵和交通事故的道路路段。因此，准确、及时地预测工作区交通事故的严重程度对于减少应急单位（如医疗援助）的响应时间至关重要，从而提高交通安全并减少拥堵。在预测交通事故的严重程度方面，以前的研究使用了不同的统计和机器学习模型，以准确性作为主要评估因素。然而，这些模型的性能通常不是很好，特别是在致命和伤害事故方面。此外，只关注预测准确性是有误导性的。本文旨在提出一种基于深度学习的新方法，该方法使用定制的 f1 损失函数来预测交通事故的严重程度。实现这一目标的基础是，考虑到两个性能指标（即精度和召回率），将深度学习模型的结果与机器学习模型进行比较。分析中使用的数据包括 2014 年至 2018 年在路易斯安那州工作区发生的交通事故样本。该数据集包括与影响这些事故发生和严重程度的道路、车辆和人为因素相关的有价值的信息（特征）。所提出的方法基于将这些特征/变量转换为图像。使用非线性降维技术 t-SNE 和凸包算法进行图像转换。使用基于 CNN 的深度学习算法和定制的损失函数直接优化模型的精度和召回率。结果表明，使用深度学习方法可以提高致命和伤害事故严重程度的预测性能，这有助于提高工作区和可能其他道路路段的交通安全和交通拥堵状况。
