Deciphering and Forecasting Highway Accidents via Realistic Synthetic Trajectories Using LLMs

Abstract: Discerning and predicting the underlying causes of road accidents involves an intricate interplay of elusive and multifaceted variables. The sporadic and often unpredictable nature of accidents poses significant challenges to researchers attempting to compile datasets that comprehensively represent the myriad scenarios that can lead to such events. Traditional analytic methodologies struggle to seamlessly integrate the complex web of contributing factors, including human behavior, environmental conditions, weather changes, and road attributes. The advent of connected vehicle technology, providing extensive spatio-temporal data through detailed travel trajectories, has introduced new opportunities for understanding road safety dynamics. However, only a minute fraction of this data directly relates to pathways that result in accidents, making it difficult to draw robust conclusions. To address these challenges, we turn to the transformative capabilities of natural language processing. Specifically, we utilize advanced Large Language Models (LLMs) to uncover latent patterns in existing data and generate synthetic yet realistic accident trajectory data. This synthetic data enriches our training datasets, serving as a foundational input for deep learning architectures focused on predicting accident likelihood based on vehicular trajectories and time-based indices. Our methodology bridges the gap created by data sparsity and surpasses the limitations of conventional statistical models, which often struggle to capture the nuanced precursors of accidents. We evaluate the efficacy of LLM-augmented machine learning models by comparing their predictions to those produced by regression-centric statistical frameworks. Initial results indicate a clear advantage in predictive performance, with LLM-integrated models achieving superior accuracy. These findings underscore the potential for integrating LLM-driven techniques into proactive road safety strategies. By enhancing predictive capabilities, such methods offer a promising pathway for developing data-driven solutions that can anticipate and mitigate accident risks, thereby contributing to safer roadways and improved public safety outcomes.