Leveraging LLMs in Cloud–Edge Networks for Traffic Risk Prediction and Accident Severity Analysis

Traffic accidents cause millions of casualties annually, making traffic safety a critical research area. Risk prediction and accident analysis are essential components of traffic safety. Traditional approaches to traffic risk prediction and accident analysis primarily rely on machine learning algorithms, which require large-scale labeled datasets and often struggle to generalize to unseen traffic conditions. Large language models (LLMs) possess strong contextual understanding and generalization capabilities, demonstrating great potential in enhancing traffic safety. LLMs have a large number of parameters, making local deployment on vehicles resource-intensive. Deploying them on the cloud or edge serves as an effective solution in cloud-edge networks. In this paper, we introduce a novel traffic safety framework leveraging LLMs in cloud-edge networks to enable risk prediction and accident severity analysis tasks. Connected autonomous vehicles (CAVs) can offload vehicle tasks to edge LLMs or cloud LLMs through cloud-edge networks, improving computational efficiency and real-time performance. This framework consists of three key modules: (1) the road segment risk prediction module, which uses LLMs to assess the risk levels (low or high) of road segments; (2) the scene risk prediction module, which utilizes LLMs to evaluate the risk levels (low, medium, or high) of real-time traffic images; and (3) the accident severity analysis module, which employs LLMs to classify the severity of accidents (minor, severe, or fatal). Experimental results demonstrate that our framework outperforms traditional machine learning methods, achieving higher accuracy in both risk prediction and accident severity analysis. This work highlights the potential of LLMs and cloud-edge networks in enhancing traffic safety.