Department of Architecture and Civil Engineering, City University of Hong Kong, Hong Kong, China.
Department of Architecture and Civil Engineering, City University of Hong Kong, Hong Kong, China.
Accid Anal Prev. 2020 Jun;141:105520. doi: 10.1016/j.aap.2020.105520. Epub 2020 Apr 8.
Traffic crash detection is a major component of intelligent transportation systems. It can explore inner relationships between traffic conditions and crash risk, prevent potential crashes, and improve road safety. However, there exist some limitations in current studies on crash detection: (1) The commonly used machine learning methods cannot simulate the evolving transitions of traffic conditions before crash occurrences; (2) Current models collected traffic data of only one temporal resolution, which cannot fully represent traffic trends in different time intervals. Therefore, this study proposes a Long short-term memory (LSTM) based framework considering traffic data of different temporal resolutions (LSTMDTR) for crash detection. LSTM is an effective deep learning method to capture the long-term dependency and dynamic transitions of pre-crash conditions. Three LSTM networks considering traffic data of different temporal resolutions are constructed, which can comprehensively indicate traffic variations in different time intervals. A fully-connected layer is used to combine the outputs of three LSTM networks, and a dropout layer is used to reduce overfitting and improve prediction performance. The LSTMDTR model is implemented on datasets of I880-N and I805-N in California, America. The results indicate that the LSTMDTR model can obtain satisfactory performance on crash detection, with the highest crash accuracy of 70.43 %. LSTMDTR models constructed on one freeway can be transferred to other similar freeways, with 65.12 % of crash accuracy on transferability. Compared with machine learning methods and LSTM models with one or two temporal resolutions, the LSTMDTR model has been validated to perform better on crash detection and transferability. A proper number of neurons in the LSTMDTR model should be determined in real applications considering acceptable detection performance and computation time. The dropout technique can reduce overfitting and improve the generalization ability of the LSTMDTR model, increasing crash accuracy from 64.49 % to 70.43 %.
交通碰撞检测是智能交通系统的重要组成部分。它可以探索交通状况与碰撞风险之间的内在关系，预防潜在的碰撞，并提高道路安全。然而，目前的碰撞检测研究存在一些局限性：（1）常用的机器学习方法无法模拟碰撞发生前交通状况的演变过程；（2）当前的模型仅收集一种时间分辨率的交通数据，无法充分代表不同时间间隔的交通趋势。因此，本研究提出了一种考虑不同时间分辨率交通数据的长短期记忆（LSTM）框架（LSTMDTR）用于碰撞检测。LSTM 是一种有效的深度学习方法，可以捕捉到碰撞前条件的长期依赖和动态变化。构建了三个考虑不同时间分辨率交通数据的 LSTM 网络，可以全面反映不同时间间隔的交通变化。使用全连接层来组合三个 LSTM 网络的输出，使用 dropout 层来减少过拟合并提高预测性能。在加利福尼亚州的 I880-N 和 I805-N 数据集上实现了 LSTMDTR 模型。结果表明，LSTMDTR 模型在碰撞检测方面具有令人满意的性能，碰撞准确率最高可达 70.43%。在可转移性方面，构建在一条高速公路上的 LSTMDTR 模型可以转移到其他类似的高速公路上，其碰撞准确率为 65.12%。与机器学习方法和具有一个或两个时间分辨率的 LSTM 模型相比，LSTMDTR 模型在碰撞检测和可转移性方面表现更好。在实际应用中，应根据可接受的检测性能和计算时间来确定 LSTMDTR 模型中适当数量的神经元。在 LSTMDTR 模型中使用 dropout 技术可以减少过拟合，提高模型的泛化能力，从而将碰撞准确率从 64.49%提高到 70.43%。
