Scholarlink

<p>Speed choice plays a major role in road accidents, yet most traffic management strategies do not fully account for differences in driver behavior or changing risk conditions. This study proposes a transparent and risk-aware framework for adaptive speed recommendation by combining unsupervised driver profiling with reinforcement learning. The analysis uses six years of road accident data from India (from 2018 to 2023) comprising tens of thousands of recorded crashes. Drivers are grouped into low, moderate and high-risk categories using Principal Component Analysis and K-means clustering. These risk profiles were then incorporated into a Proximal Policy Optimization agent trained with a safety-focused reward function that reflects driver behavior, road characteristics, environmental conditions and accident severity. The learned policy consistently recommends lower speeds for high-risk drivers and adverse conditions, while allowing higher yet safe speeds in low-risk contexts. Offline proxy-based evaluation using historical accident data shows that moderate speeds in the range of about 50–60&#xa0;km/h were associated with lower average fatalities and casualties. Model interpretability is supported through SHAP analysis and decision-tree approximations which identify driver age, risk category and posted speed limits as key factors influencing recommendations. Although the framework is evaluated using offline and correlational analysis rather than field deployment, the results demonstrate the potential of interpretable and risk-aware AI systems to support safer speed management and inform intelligent transportation policy decisions.</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

Adaptive and Interpretable Speed Control for Accident Prevention Using Driver Risk Profiling

  • 期刊论文

  • 发表日期：

    2026年4月30日

查看全文
  • Priyam Nath Bhowmik,
  • Kezia Saini,
  • Boya Veeresh,
  • N. Sivabalaji,
  • V. Sai Kumar

摘要

Speed choice plays a major role in road accidents, yet most traffic management strategies do not fully account for differences in driver behavior or changing risk conditions. This study proposes a transparent and risk-aware framework for adaptive speed recommendation by combining unsupervised driver profiling with reinforcement learning. The analysis uses six years of road accident data from India (from 2018 to 2023) comprising tens of thousands of recorded crashes. Drivers are grouped into low, moderate and high-risk categories using Principal Component Analysis and K-means clustering. These risk profiles were then incorporated into a Proximal Policy Optimization agent trained with a safety-focused reward function that reflects driver behavior, road characteristics, environmental conditions and accident severity. The learned policy consistently recommends lower speeds for high-risk drivers and adverse conditions, while allowing higher yet safe speeds in low-risk contexts. Offline proxy-based evaluation using historical accident data shows that moderate speeds in the range of about 50–60 km/h were associated with lower average fatalities and casualties. Model interpretability is supported through SHAP analysis and decision-tree approximations which identify driver age, risk category and posted speed limits as key factors influencing recommendations. Although the framework is evaluated using offline and correlational analysis rather than field deployment, the results demonstrate the potential of interpretable and risk-aware AI systems to support safer speed management and inform intelligent transportation policy decisions.