Understand the impact of traffic states on crash risk in the vicinities of Type A weaving segments: A deep learning approach.

The primary objective of this study was to evaluate the impacts of traffic states on crash risk in the vicinities of Type A weaving segments. A deep convolutional embedded clustering (DCEC) was developed to classify traffic flow into nine states. The proposed DCEC outperformed the three common clustering algorithms, i.e. K-means, deep embedded clustering, and deep convolutional autoencoders clustering, in terms of silhouette coefficient and calinski-harabaz index on the same samples, suggesting that the DCEC provides better clustering performance. The characteristics of the nine traffic states are described for the right and inside lanes separately. The DCED visualization indicates that the spatiotemporal features of the nine traffic states are different from each other. The empirical analyses suggest that crash severity and the main types of crashes are different across the nine traffic states. The results of the logistic regression model prove that the nine traffic states are significantly associated with crash risk in the vicinities of weaving segments, and each traffic state can be assigned with a unique safety level. The convolutional neural network with gated convolutional layers (G-CNN) was developed to predict the crash risk in each traffic state. Compared with the traditional four traffic states classification based on 4-phase traffic theory, the model incorporating the various crash mechanisms across the nine traffic states provides more accurate predictions.