LLM Multimodal Traffic Accident Forecasting.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Urban Planning

Background:

• Increasing urban traffic congestion necessitates advanced accident prediction for public safety.
• Current autonomous driving systems require enhanced real-time decision-making capabilities.

Purpose of the Study:

• To evaluate deep learning (DL) models for traffic accident forecasting.
• To integrate large language models (LLMs) and multimodal AI into Level-4/Level-5 autonomous driving systems.
• To improve real-time responsiveness and feature importance analysis in self-driving scenarios.

Main Methods:

• Comparative analysis of Transformer models against ARIMA and Prophet for accident prediction.
• Principal Component Analysis (PCA) for feature importance identification.
• Integration of lightweight LLMs (LLaMA-2, Zephyr-7b-α) and Large Language-and-Vision Assistant (LLaVA) with deep probabilistic reasoning.

Main Results:

• Deep learning models, particularly Transformers, demonstrated superior performance in traffic accident forecasting compared to traditional time series models.
• Identified key accident contributing factors through PCA.
• Successfully integrated multimodal AI and LLMs to enhance autonomous driving system responsiveness.

Conclusions:

• Deep learning and multimodal AI, including LLMs, significantly advance traffic accident prediction and autonomous driving safety.
• Data-driven insights from DL and probabilistic programming are crucial for developing smarter, safer urban environments.
• The study highlights the potential of advanced AI for real-time interventions in autonomous vehicles.