Intelligent Driving Collision Risk Warning Based on Fine-Tuned Large Language Models

• 1、School of Computer Science, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Collision risk warning systems enhance driving safety by monitoring and analyzing the driving environment to issue timely alerts before potential accidents occur. To address the limitations of existing collision risk assessment methods-specifically the low accuracy in identifying collision objects, restricted risk coverage, static warning levels, and the insufficient inference efficiency of Large Language Model (LLM) approaches-this paper proposes an intelligent driving collision risk warning method based on fine-tuned LLMs. First, a potential field theory-based assessment is employed to pre-screen perception data, integrating factors such as mass, velocity, and orientation into a unified potential field framework. Second, Quantized Low-Rank Adaptation is applied to fine-tune pre-trained LLMs, enabling them to better comprehend collision scenarios and analyze potential collision targets as well as Time-to-Collision (TTC). Subsequently, a risk level assessment method calculates a comprehensive risk score by assigning specific weights to factors including visibility, temporal variables, road surface slipperiness, speeding risks, and TTC. Finally, a structured risk warning workflow is established to facilitate the full process from scenario data input to collision risk warning output. Experimental results demonstrate that the proposed method outperforms both conventional potential field methods and existing LLM-based solutions, while also verifying its consistent effectiveness across various LLM architectures.

Keywords： Computer Applied Technology Intelligent Driving Collision Risk Warning Large Language Models