Behavioral fingerprints: driver profiling using transformer models on next generation simulation trajectory data

Abstract

Characterizing individual driver behavior is essential for advancing intelligent transportation systems (ITS) and autonomous vehicle safety. While deep learn ing models excel at macroscopic traffic prediction, individual driving styles are often aggregated away. This paper addresses this gap by proposing a novel, weakly supervised transformer framework for driver behavior profiling using high-resolution next generation simulation (NGSIM) US-101 trajectory data. We extract microscopic behavioral features including acceleration, lane change dynamics, and headway management from 30-second observation segments. A transformer encoder learns complex temporal dependencies to classify drivers into ’aggressive’ and ’normal’ profiles, achieving a 97% F1-score on proxy labeled segments. Crucially, these “proxy labels” are derived from heuristic statistics, meaning the model is trained to learn the mapping from sequences to these behavioral indicators rather than identifying objective aggression. Our methodology enables the creation of precise “behavioral fingerprints” that cap ture individual driving nuances. These insights are vital for developing adaptive ITS that anticipate traffic stability issues and enhance autonomous vehicle safety by predicting human intent.