In this paper, a gated recurrent unit–deep neural network (GRU-DNN) model integrated with multimodal feature embedding (MFE) is developed to evaluate the real-time risk of hazmat road transportation based on various types of data for contributing factors. MFE was incorporated into the framework of a deep learning model in which discrete variables, continuous variables, and images were uniformly embedded. GRU is a pre-trained sub-model, and the DNN is able to directly use the relative structure and weights of the GRU, improving the poor classification and recognition results due to insufficient samples. Additionally, the model is trained and validated based on hazmat road transportation database consisting of 2100 samples with 20 real-time contributing factors and four risk levels in China. The accuracy (ACC), precision (PR), recall (RE), F1-score (F1), and areas under receiver-operating-characteristic curves (AUC) of the proposed model and other commonly used models are compared as performance measurements in numerical examples. Finally, Carlini & Wagner attack and three defenses of adversarial training, dimensionality reduction and prediction similarity are proposed in the training to improve the robustness of the model, alleviating the impact of noise and error on small-sized samples. The results demonstrate that the average ACC of the model reaches 93.51% and 87.6% on the training and validation sets, respectively. The prediction of accidents resulting in injury is the most accurate, followed by fatal accidents. Combined with the RE of 89.0%, the model exhibits excellent performance. In addition, the proposed model outperforms other widely used models based on the overall comparisons of ACC, AUC, F1 and PR-RE curve. Finally, prediction similarity can be used as an effective approach for robustness improvement, with the launched adversarial attacks being detected at a high success rate.

Nowadays, conditional automated driving vehicles still need drivers to take-over in the scenarios such as emergency hazard events or driving environments beyond the system’s control. This study aimed to explore the changing trend of the drivers’ takeover behavior under the influence of traffic density and take-over budget time for the entire take-over process in emergency obstacle avoidance scenarios. In the driving simulator, a 2 × 2 factorial design was adopted, including two traffic densities (high density and low density) and two kinds of take-over budget time (3 s and 5 s). A total of 40 drivers were recruited, and each driver was required to complete four simulation experiments. The driver’s take-over process was divided into three phases, including the reaction phase, control phase, and recovery phase. Time parameters, dynamics parameters, and operation parameters were collected for each take-over phase in different obstacle avoidance scenarios. This study analyzed the variability of traffic density and take-over budget time with take-over time, lateral behavior, and longitudinal behavior. The results showed that in the reaction phase, the driver’s reaction time became shorter as the scenario urgency increased. In the control phase, the steering wheel reversal rate, lateral deviation rate, braking rate, average speed, and takeover time were significantly different at different urgency levels. In the recovery phase, the average speed, accelerating rate, and take-over time differed significantly at different urgency levels. For the entire take-over process, the entire take-over time increased with the increase in urgency. The lateral take-over behavior tended to be aggressive first and then became defensive, and the longitudinal take-over behavior was defensive with the increase in urgency. The findings will provide theoretical and methodological support for the improvement of take-over behavior assistance in emergency take-over scenarios. It will also be helpful to optimize the human-machine interaction system.

This paper models and mitigates the secondary crash (SC) risk for serial tunnels on the freeway which is incurred by traffic turbulence after primary crash (PC) occurrence and location-heterogeneous lighting conditions along serial tunnels. A traffic conflict approach is developed where SC risk is quantified using a surrogate safety measure based on the simulated vehicle trajectories after PC occurs from a lighting-related microscopic traffic model with inter-lane dependency. Numerical examples are presented to validate the model, illustrate SC risk pattern over time, and evaluate the countermeasures for SC, including adaptive tunnel lighting control (ATLC) and advanced speed and lane-changing guidance (ASLG) for connected vehicles (CVs). The results demonstrate that the tail of the stretching queue on the PC occurrence lane, the adjacent lane of the PC-incurred queue, and areas near tunnel portals are high-risk locations. In serial tunnels, creating a good lighting condition for drivers is more effective than advanced warnings in CVs to mitigate SC risk. Combined ATLC and ASLG is promising since ASLG informs CVs of an immediate response to traffic turbulence on the lane where PC occurs and ATLC alleviates SC risks on adjacent lanes via smoothing the lighting condition variations and reducing inter-lane dependency.