Applying Few-Shot Learning in Classifying Pedestrian Crash Typing

Abstract

Pedestrian deaths account for 23% of all road traffic fatalities worldwide. After declining for three decades, pedestrian fatalities in the United States have been increasing with 6,941 fatalities in 2020, the highest number for more than two decades, impeding progress toward a zero-deaths transportation system. The Pedestrian and Bicycle Crash Analysis Tool (PBCAT) was developed to describe the pre-crash actions of the parties involved to better define the sequence of events and precipitating actions that lead to crashes involving motor vehicles and pedestrians or cyclists. Undoubtedly, police crash data influence decision-making processes in the transportation agencies. Using crash data from three major cities in Texas (during the period from 2018 to 2020), this study assessed the data quality of police-reported crash narratives on pedestrian-involved traffic crashes. As the pedestrian crash typing involves many categories, conventional machine-learning algorithms will not be sufficient in solving the classification problem from narrative texts. This study used few-shot learning (FSL), an advanced machine learning, to solve this issue. Using the pre-knowledge obtained from five different crash types and a few labeled data points of three unseen new crash types, the proposed model achieved roughly 40% overall accuracy. Also, four different configurations of crash types were formed and tested which indicates that the model is robust.