Scholarlink

<p>Road-based traffic accident prediction aims to prevent accidents by analyzing historical accident data and related road factors. However, existing methods overlook the spatial diffusion of accident impact, nonlinear volatility, and comprehensive spatial correlations. To address these limitations, this paper proposes a Traffic Accident Prediction Model Based on Diffusion Modeling and Spatiotemporal Dependency Awareness (DMSDAnet), formulating prediction as a continuous risk regression problem. The model incorporates directed graphs with an enhanced Graph Attention Network (GAT) to capture risk diffusion, utilizes a linear ordinary differential equation (ODE) to model dynamic volatility, and integrates road attributes to enhance spatial correlations. Validated on the Sacramento dataset, DMSDAnet outperforms eight baselines, improving Recall and RMSE by 3.45% and 3.68%, respectively. Additionally, the architecture is optimized for high-performance computing (HPC), enabling real-time inference on large-scale networks suitable for latency-critical Intelligent Transportation Systems (ITS).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Traffic accident prediction model based on diffusion modeling and spatiotemporal dependency awareness

  • 期刊论文

  • 发表日期：

    2026年3月13日

查看全文
  • Jian Feng,
  • Kun Qian,
  • Yue Wang

摘要

Road-based traffic accident prediction aims to prevent accidents by analyzing historical accident data and related road factors. However, existing methods overlook the spatial diffusion of accident impact, nonlinear volatility, and comprehensive spatial correlations. To address these limitations, this paper proposes a Traffic Accident Prediction Model Based on Diffusion Modeling and Spatiotemporal Dependency Awareness (DMSDAnet), formulating prediction as a continuous risk regression problem. The model incorporates directed graphs with an enhanced Graph Attention Network (GAT) to capture risk diffusion, utilizes a linear ordinary differential equation (ODE) to model dynamic volatility, and integrates road attributes to enhance spatial correlations. Validated on the Sacramento dataset, DMSDAnet outperforms eight baselines, improving Recall and RMSE by 3.45% and 3.68%, respectively. Additionally, the architecture is optimized for high-performance computing (HPC), enabling real-time inference on large-scale networks suitable for latency-critical Intelligent Transportation Systems (ITS).