A Bayesian extreme value theory modelling framework to assess corridor-wide pedestrian safety using autonomous...

Area of Science:

• Transportation Engineering
• Road Safety Analysis
• Data Science

Background:

• Pedestrian safety is critical, but crashes are dispersed, making corridor-level risk assessment challenging.
• Existing studies often focus on intersections, lacking network-wide pedestrian interaction models.
• Autonomous vehicle data offers a rich, untapped resource for comprehensive road safety analysis.

Purpose of the Study:

• To develop and apply an extreme value theory (EVT) framework for estimating corridor-wide pedestrian crash risk.
• To utilize autonomous vehicle (AV) sensor data for network-level safety analysis.
• To compare the performance of different EVT models in predicting pedestrian crash frequencies.

Main Methods:

• Developed Bayesian extreme value theory models: Generalised Extreme Value (GEV) for block maxima and Generalised Pareto (GPD) for peak over threshold.
• Applied the framework to AV sensor data (LiDAR, cameras) from the Argoverse dataset in Miami, USA.
• Extracted vehicle-pedestrian trajectories, identified conflicts using Post Encroachment Time (PET), and incorporated traffic/pedestrian volumes and speeds as covariates.

Main Results:

• Both GEV and GPD models provided reasonable estimates of historical pedestrian crash frequencies.
• The block maxima (GEV) model demonstrated higher accuracy than the peak over threshold (GPD) model.
• Non-stationarity was effectively captured by including traffic and pedestrian flow variables.

Conclusions:

• Autonomous vehicle sensor data is a valuable resource for network-level pedestrian safety analysis.
• The proposed EVT framework enables efficient identification of pedestrian crash risk zones across transport networks.
• This approach advances the understanding and mitigation of pedestrian risks beyond isolated intersections.