Semantic modelling of pedestrian behaviours at shared-use path crossings: insights from BERT-based architectures and structural topic modelling

View at publisher Summarise with VIRAT
Authors: Mahmuda Sultana Mimi, Md Monzurul Islam, Subasish Das, Anandi K Dutta, Bahar Dadashova
Cited by: 0
References: 135

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Pedestrian safety at urban shared-use path (SUP) crossings is a growing concern in expanding metros, where conflicts occur between motorised and non-motorised users. This study examines pedestrian crossing behaviours at SUP intersections in five Texas cities, Austin, Houston, Dallas, San Antonio, and El Paso, where high urban density and multimodal traffic increase crash risk. Traditional tools such as the Pedestrian and Bicycle Crash Analysis Tool (PBCAT) rely on manual classification, leading to inefficiencies and inconsistencies. To address this, an automated semantic framework using Transformer-based models (BERT, RoBERTa, DistilBERT) is developed to classify pedestrian maneuvers based on PBCAT definitions. Structural Topic Modeling (STM) is applied to crash narratives (2020–2023) to identify latent themes by maneuver type. To mitigate class imbalance, synthetic narratives are generated using Large Language Models. Models are fine-tuned and evaluated using F1-score and AUC. Results show BERT performs best, while STM highlights key risk factors, informing targeted safety interventions.

Fields of Study & Subjects

Traffic and Road Safety Urban Transport and Accessibility Evacuation and Crowd Dynamics

Publication details

Journal / source
Transportmetrica A Transport Science
Pages
1–53
Publication date
30 March 2026
Language
EN