Recent developments in artificial intelligence (AI) have made significant improvements in understanding and enhancing pedestrian safety—a vulnerable road user group that receives less attention than motorized road users do. Specifically, AI-based video analytics have provided insight into facilitating real-time safety at signalized intersections. However, past studies have not fully realized the essence of real-time analysis, which underpins forecasting pedestrian collision likelihood by analyzing how past extreme events influence future risk over sequential intervals. To this end, we combine extreme value theory and machine learning models for real-time pedestrian collision risk forecasting. Traffic conflicts and their associated variables were identified from 288 ​h of video footage obtained from three signalized intersections in Queensland, Australia, via computer vision techniques, including YOLO and DeepSORT, to obtain the post encroachment time for vehicle‒pedestrian interactions. A Bayesian non-stationary peak over threshold (POT) is developed to obtain real-time pedestrian crash risk at the signal cycle level. The performance of the POT model is compared with observed crashes, and the results demonstrate the reasonable accuracy of the model. The estimated pedestrian crash risk at each signal cycle forms contiguous univariate time series data (which serve as ground truth), which are used as input to develop time series machine learning models (recurrent neural networks (RNNs) and long short-term memory (LSTM)). Both of these models forecast pedestrian crash risk, with the RNN model outperforming the competing model and demonstrating that pedestrian crash risk can be reliably estimated 30−33 ​min in advance.

Traffic conflict techniques rely heavily on the proper identification of conflict extremes, which directly affects the prediction performance of extreme value models. Two sampling techniques, namely, block maxima and peak over threshold, form the core of these models. Several studies have demonstrated the inefficacy of extreme value models based on these sampling approaches, as their crash estimates are too imprecise, hindering their widespread practical use. Recently, anomaly detection techniques for sampling conflict extremes have been used, but their application has been limited to estimating crash frequency without considering the crash severity aspect. To address this research gap, this study proposes a hybrid model of machine learning and extreme value theory within a bivariate framework of traffic conflict measures to estimate crash frequency by severity level. In particular, modified time-to-collision (MTTC) and expected post-collision change in velocity (Delta-V or ΔV) have been proposed in the hybrid modeling framework to estimate rear-end crash frequency by severity level. Rear-end conflicts were identified through artificial intelligence-based video analytics for three four-legged signalized intersections in Brisbane, Australia, using four days of data. Non-stationary bivariate hybrid generalized extreme value models with different anomaly detection/sampling techniques (isolation forest and minimum covariance determinant) were developed. The non-stationarity of traffic conflict extremes was handled by parameterizing model parameters, including location, scale, and both location and scale parameters simultaneously. The results indicate that the bivariate hybrid models can estimate severe and non-severe crashes when compared with historical crash records, thereby demonstrating the viability of the proposed approach. A comparative analysis of two anomaly techniques reveals that the isolation forest model marginally outperforms the minimum covariance determinant model. Overall, the modeling framework presented in this study advances conflict-based safety assessment, where the severity dimension can be captured via bivariate hybrid models.

Lane-changing is performed either to follow the route to a planned destination (i.e., mandatory lane-changing) or to achieve better driving conditions (i.e., discretionary lane-changing). A connected environment is expected to assist during lane-changing manoeuvres, but it is not known well how driving aids in a connected environment assist lane-changing execution. As such, this study investigates the impact of a connected environment on lane-changing execution time during mandatory and discretionary lane-changing manoeuvres. To this end, this study designed an advanced driving simulator experiment where 78 drivers performed these manoeuvres on a simulated motorway in three randomised driving conditions. The conditions were baseline (without driving aids), a fully functioning connected environment with a perfect supply of driving aids, and an impaired connected environment with delayed driving aids. The lane-changing execution time has been modelled by a random parameters hazard-based duration modelling approach, which accounts for the panel nature of data and captures the unobserved heterogeneity. Results suggest that, compared to the baseline condition (i.e., a non-connected environment), most of the drivers in the connected environment take more time to complete their lane-changing manoeuvres, indicating drivers’ safer lane-changing execution behaviour in the connected environment. The communication delay driving condition has been found to have more deteriorating effects on mandatory lane-changing manoeuvres than discretionary lane-changing manoeuvres. This study concludes that (i) the connected environment increases safety margin during both lane-changing manoeuvres, and (ii) a higher magnitude of safety margin is observed during mandatory lane-changing manoeuvres whereby drivers have a higher need for assistance.