Traffic safety has always been an important issue in sustainable transportation development, and the prediction of traffic accident severity remains a crucial challenging issue in the domain of traffic safety. A huge variety of forecasting models have been proposed to meet this challenge. These models gradually evolved from linear to nonlinear forms and from traditional statistical regression models to current popular machine learning models. Recently, a machine learning algorithm called Deep Forests based on the decision tree ensemble has aroused widespread concern, which was proposed for the first time by a research team of Nanjing University. This algorithm was proved to be more accurate and robust in comparison with other machine learning algorithms. Motivated by this benefit, this study employs the UK road safety dataset to propose a novel method for predicting the severity of traffic accidents based on the Deep Forests algorithm. To verify the superiority of our proposed method, several other machine learning algorithm-based perdition models were implemented to predict traffic accident severity with the same dataset, and the prediction results show that the Deep Forests algorithm present good stability, fewer hyper-parameters, and the highest accuracy under different level of training data volume. It is expected that the findings from this study would be helpful for the establishment or improvement of effective traffic safety system within a sustainable transportation system, which is of great significance for helping government managers to establish timely proactive strategies in traffic accident prevention and effectively improve road traffic safety.

This paper evaluates the traffic safety of freeway interchange merging areas based on the traffic conflict technique. The hourly composite risk indexes (HCRI) was defined. By the use of unmanned aerial vehicle (UAV) photography and video processing techniques, the conflict type and severity was judged. Time to collision (TTC) was determined with the traffic conflict evaluation index. Then, the TTC severity threshold was determined. Quantizing the weight of the conflict by direct losses of different severities of freeway traffic accidents, the calculated weight of the HCRI can be obtained. Calibration of the relevant parameters of the micro-simulation simulator VISSIM is conducted by the travel time according to the field data. Variables are placed into orthogonal tables at different levels. On the basis of this table, the trajectory file of every traffic condition is simulated, and then submitted into a surrogate safety assessment model (SSAM), identifying the number of hourly traffic conflicts in the merging area, a statistic of HCRI. Moreover, the multivariate linear regression model was presented and validated to study the relationship between HCRI and the influencing variables. A comparison between the HCRI model and the hourly conflicts ratio (HCR), without weight, shows that the HCRI model fitting degree was obviously higher than the HCR. This will be a reference to design and implement operational planners.

Accurate vehicle trajectory prediction is critical for autonomous vehicles and advanced driver assistance systems to make driving decisions and improve traffic safety. This paper proposes a novel Temporal Multi-task Mixture Of Experts (TMMOE) model for simultaneously predicting the vehicle trajectory and driving intention, considering interconnections among tasks. As for the methodology, the proposed model consists of three layers: a shared layer, an expert layer, and a fully connected layer. In more detail, the first layer utilizes Temporal Convolutional Network (TCN) to extract temporal features whereas the expert layer incorporates the gating mechanism to memorize and filter the temporal dependence of sequences and, finally, the fully connected layer is applied to integrate and export prediction results. Furthermore, the homoscedastic uncertainty algorithm is used to construct the multi-task loss function. The open data source CitySim dataset is chosen to validate the performance of the proposed TMMOE model; moreover, a novel lane line reconstruction method is introduced to mitigate measurement errors of the dataset. Two-part information, including the history of the vehicle's trajectory and the interaction indicators, are employed as the input variables. The result indicates that the TMMOE algorithm exhibits superior performance when compared to the other five models, including Long Short-Term Memory (LSTM) Network, Convolutional Neural Network (CNN), CNN-LSTM, TCN, and Gate Recurrent Unit (GRU), while considering varying input sequence lengths are 3 s, 6 s, and 9 s respectively. Finally, the sensitivity analysis demonstrates that considering driving intentions in vehicle trajectory prediction can significantly enhance the prediction accuracy of vehicle trajectories.

The severity of the two-vehicle crash is closely related to the characteristics of both the struck and striking vehicles. Ignoring vehicle roles may lead to biased results. Thus, this study used mixed logit models to determine the factors that influence injury severity in the two-vehicle crash, taking into account the vehicle characteristics of the different crash roles. The data used is collected from Pennsylvania Department of Transportation (PennDOT) Open Data Portal. First, the synthetic minority oversampling technique and nearest neighbors (SMOTE-ENN) strategy was selected to address the class imbalance problem of crash data. Then, two separated mixed logit models were developed for four- and three-legged unsignalized intersections. The results suggest that the type and movement of vehicles have significant effects on crash severity. For example, right-turn vehicles being struck can lead to more serious crashes than striking other vehicles. Large trucks striking other vehicles are found to increase crash severity, but being struck is found to decrease crash severity. Additionally, several factors were also identified to affect crash severity in both models and effective countermeasures suggestions were proposed to mitigate crash severity. Supplemental data for this article is available online at at .