Predicting pedestrian-vehicle interaction severity at unsignalized intersections.

Area of Science:

• Computer Science
• Transportation Engineering
• Artificial Intelligence

Background:

• Pedestrian-vehicle interactions at unsignalized intersections pose significant safety risks.
• Traditional models often lack the interpretability and accuracy needed for complex dynamic environments.
• Developing advanced predictive models is crucial for enhancing road safety.

Purpose of the Study:

• To develop and validate a novel deep-learning model for predicting pedestrian-vehicle interaction severity.
• To integrate Transformer-based models with Multilayer Perceptrons (MLP) for enhanced feature analysis.
• To provide a more interpretable prediction method compared to traditional approaches.

Main Methods:

• Collected high-resolution optical camera data of pedestrian and vehicle movements.
• Extracted trajectories and georeferenced them for precise spatial analysis.
• Developed a hybrid Transformer-MLP deep-learning model trained on dynamic variables, utilizing attention mechanisms for interpretability.

Main Results:

• The model achieved high performance, with an overall accuracy of 0.87.
• Precision, recall, and F1-scores exceeded 0.84 for safe interactions and were higher for critical events and conflicts.
• Variable importance analysis identified 'Vehicle Speed' as a key positive influencer, while 'Approaching Angle' and 'Vehicle Distance' were negative influencers.

Conclusions:

• A novel Transformer-MLP deep-learning model effectively predicts pedestrian-vehicle interaction severity at crosswalks.
• The model demonstrates high precision and recall, offering valuable insights into safety-critical factors.
• Findings support advancements in real-time safety analysis and urban traffic safety assessments.