Stress is considered by many studies to affect traffic safety, and many researchers have attempted to monitor the dynamics of driving stress. Previous research has relied excessively on the positive effects of psychological indicators to improve the accuracy of stress monitoring models. However, psychological data collection sensors have not been widely used in conventional vehicles, which makes it impossible to apply the results of that research to actual driving tasks on a daily basis, even if the accuracy is high. This study designs a real driving task to extract data and proposes a driver's driving stress monitoring model based on driving behaviour, driving environment, and route familiarity. The driving behaviour is described by the speed and acceleration of the vehicle, and the driving environment is quantified by a dilated residual networks (DRN) model thazt divides the video image from the full region into subregions according to the distribution of the driver's attention. Based on the psychological data and driver stress inventory (DSI) results, the study used a K-means 3D cluster analysis to obtain the evaluation method of driving stress and constructed an extreme gradient boosting (XGBoost) model to monitor driving stress. Comparisons of performance with other models show that the XGBoost model significantly outperforms the other three mainstream machine learning algorithms and exceeds most traditional models without the use of psychological data. The model's performance indicators, accuracy, sensitivity, and precision, reached 91.18%-93.25%, 84.13%-89.37%, and 90.25%-91.34%, respectively. The study also summarises the ranking of effects of different scene elements on driving stress for each visual field. The results could make it possible to apply stress monitoring on a large scale to real driving situations, providing urban designers with advice on how to reduce driver stress and directing their attention to those visual areas and visual scene elements that have a higher impact on driving stress and need improvement.

Highway tunnels have a higher risk of crashing than open roads, which require a systematic approach to tunnel safety. However, previous research had the following problems: (1) Studies have largely focused on open roads, with very little research on tunnels. (2) The collected crash contributing factors involve narrow ranges, with very little tunnel crash data including both tunnel design features, traffic conditions and pavement conditions. (3) None of the studies considered both excess zero observations and unobserved heterogeneity with its interactions. To address these issues, this paper first established an appropriate tunnel dataset containing 3 to 5 years of crash data from several highways in China and the influence factors of tunnel design features, traffic conditions and pavement conditions. A correlated random parameters negative binomial Lindley (CRPNB-L) model that considers both excess zero observations and unobserved heterogeneity with its interaction effects was then proposed. Compared to the uncorrelated random parameters negative binomial Lindley (URPNB-L) model, fixed parameters negative binomial Lindley (FPNB-L) model and fixed parameters negative binomial (FPNB) model, the CRPNB-L model solves the deviation that arises from excess zero observations by introducing the Lindley distribution and considers the unobserved heterogeneity with its interactions by introducing correlated random parameters. In the comparisons, the CRPNB-L model achieves the best effects in the goodness-of-fit. Furthermore, the estimated results of the CRPNB-L model showed that segment length, traffic volume, proportion of class 5 vehicle (heavy trucks and trailers), tunnel entrance and exit segments, and steep uphill and downhill segments were associated with higher crash frequency, while curvature, tunnel length, pavement damage condition index (PCI) and skid resistance index (SRI) were associated with lower crash frequency. In addition, the random variables of the curvature, the steep downgrade indicator, the proportion of class 5 vehicle and SRI were identified and their intercorrelations were analyzed.

Driving stress is the demand for reserved cognitive space after a driver perceives changes in vehicle, road and environmental factors during driving, which has been proven to affect driving behaviour, interfering with driving safety. Traditional stress prediction relies extensively on psychological data and is limited by the unpopularity of psychological data collection technology, which cannot be applied in daily life on a large scale. In recent years, advances in high-precision visual analysis technology represented by deep learning have laid the foundation for automated and large-scale visual environment analysis. This study proposes a framework for the quantitative analysis of highway driving stress based on multiple vehicle, road, and environmental factors. A dilated residual network model and other methods were used to extract visual environmental indexes. Combined with multisource data such as traffic volume and road design parameters, the LightGBM method was used to construct an expressway driving stress prediction model with high accuracy. The MAE, RMSE and $R^{2}$ values of the proposed model are 0.042, 0.004 and 0.881, respectively, demonstrating the usefulness for scaled and efficient assessment of expressway stress loads. The SHAP method was used to explore the relationship between different influencing factors and driving stress to quantify the mechanism of vehicle, road and environment influences on stress load, and to propose recommendations for highway design and planning from the perspective of reducing stress load. This study provides a new way of thinking to quantitatively investigate the link between multiple road traffic factors and driving stress, providing efficient and large-scale assessment of expressway driving stress, as well as proposing some suggestions for highway design and planning to enhance stress reduction.

The existing crash modeling techniques for expressway tunnels must overcome the following difficulties: 1) The collected risk factors contributing to the tunnel crashes include narrow ranges, especially the pavement conditions and weather conditions of the tunnels are rarely taken into account. 2) Most researchers ignored the estimation deviation caused by the excess zero observations of tunnel crash datasets. 3) No existing tunnel crash model can combine the random-parameters approach and spatial-temporal approach to solve the estimation deviation caused by the inter-samples and spatial-temporal heterogeneity. To address these problems, this study presents an investigation of the safety effects of risk factors of tunnel design features, traffic conditions, pavement conditions and weather conditions utilizing a 12-quarter period (3 years) of data as well as five crash frequency models: 1) a fixed parameters negative binomial model (FPNB), 2) a random parameters negative binomial model (RPNB), 3) a random parameters negative binomial Lindley model (RPNBL), 4) a spatial and random parameters negative binomial Lindley model (SP-RPNBL), and 5) a spatial-temporal and random parameters negative binomial Lindley model (ST-RPNBL). The results showed that the ST-RPNBL model solves the deviation that arises from excess zero observations by introducing the Lindley distribution and considers the unobserved heterogeneity by introducing both the random parameters and spatial-temporal parameters that provided better goodness of fit and offered more insights into the factors that contribute to tunnel safety. Furthermore, the ST-RPNBL model detected 16 variables that were significantly correlated with tunnel crash frequency, of which 12 variables were associated with a higher crash frequency and four variables were associated with a lower crash frequency. The random variables of the curvature, the steep downgrade indicator, the proportion of class 5 vehicle and the skidding resistance index (SRI) were identified, and the influence of each significant variable on the crash frequency was analyzed.

As hazardous locations of a road, freeway tunnels have a higher risk of casualty than open roads. Therefore, it is necessary to seek a reliable crash prediction model and propose targeted improvement measures. However, existing studies on freeway tunnel crash models mainly suffer from the following problems: 1) They ignore the correlation between different injury severity levels of crashes; 2) They ignore the impact of excess zero observations; 3) They do not consider the influence of heterogeneity between samples and the spatio-temporal correlation. To solve the above problems, this paper has compiled a dataset with freeway tunnel design features, three years of traffic conditions, pavement conditions, and traffic crash data. Then, a bivariate random parameters negative binomial Lindley model (ST-BRPNB-L) is established for jointly modeling crash counts and injury severity levels, which consider excess zero observations by introducing Lindley parameters, characterize the heterogeneity, and spatial-temporal correlation between samples by introducing random parameters and spatio-temporal parameters. The Bayesian estimation results have shown that ST-BRPNB-L has the best goodness-of-fit among a series of comparison models, which verifies the superiority of the proposed model. On this basis, the influence of the risk factors on the frequency and severity of crashes was quantitatively analyzed based on the ST-BRPNB-L model’s parameters estimation results, which provides a scientific basis for safety improvement measures of freeway tunnels.

Although a substantial number of traffic videos have been accumulated via daily monitoring, deep learning is seldom utilized to process these data for multilevel traffic state detection. The application of deep learning is limited for two reasons: (1) the multilevel traffic state based on traffic images has not been defined. (2) The high noise information in traffic images and extremely similar features of adjacent traffic states hinder accurate detection. Based on this situation, A new definition of the image-based multilevel traffic state is proposed using the ratio of the vehicle areas to the road areas in a traffic image, and a standard image dataset, including various illuminations and vast scenes, are established. A deep residual network named TrafficNet, which is embedded with Squeeze-and-Excitation blocks and is learned by the improved triplet loss, is proposed for multilevel traffic state detection. The Squeeze-and-Excitation block effectively reduces the model’s attention to noise information and focuses on road areas that are associated with traffic features in an image. The improved triplet loss maps the learned features to a metric space where the distance between features of inter-class is larger than that within the same class, which improves the discrimination of features between adjacent traffic states. Relevant experiments prove that the performance of TrafficNet, whose accuracy (Acc) in classifying 10 traffic states reaches 94.27% with the testing dataset, is much better than that of traditional deep classification models, which do not include Squeeze-and-Excitation blocks or the improved triplet loss.