Classifying Pedestrian Maneuver Types Using the Advanced Language Model

Abstract

Pedestrians are the most vulnerable roadway user. While there is much emphasis on “green transportation,” a troubling fact emerges in the U.S.A.: pedestrian deaths are increasing significantly in comparison to motorist deaths, reaching nearly 6941 in 2020—the highest in over two decades. The Pedestrian and Bicycle Crash Analysis Tool was developed to determine motorists and non-motorists’ actions before a crash to accurately define the sequence of events and precipitating actions leading to traffic crashes between motor vehicles and pedestrians or bicyclists. Police report traffic crash data and crash narrative reports undoubtedly play a major role in decision-making for the safety engineers. Using crash data from three major cities in Texas (2018–2020), this study assessed the data quality of text narratives in police reports of pedestrian crashes. The objective of this study was to develop a framework for applying advanced language models to classify pedestrian maneuver types from unstructured textual content. The results show that although natural language processing models are promising as crash typing tools, narration inconsistency, data imbalance, and small sample sizes are holding back progress in this area. The framework demonstrated high accuracy for the binary classification task, but it was inconsistent for the more complex multiclass task. This framework provides the basis for applying advanced language models such as the bidirectional encoder representations from transformers model in identifying pedestrian maneuver types associated with pedestrian crashes.