<b>K-Risk:</b> A knowledge-augmented dataset of high-risk driving scenarios with LLM annotations for autonomous driving

Engineering notes

K-Risk aggregates <b>20 human-driven (HV) and automated-vehicle (AV) trajectory sources</b> across Europe, China and the United States — covering highways, urban freeways, intersections and roundabouts — and curates <b>31,398 high-risk events</b>, including a <b>1,036-event extreme near-collision subset</b>.The <b>data-processing pipeline</b> (event extraction, scenario-description generators, the closed-loop LLM annotation driver, and the statistics/figure scripts) is distributed separately as source code; see https://github.com/benmagnifico/K-Risk.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

This archive is the <b>data release</b> of K-Risk. It pairs <b>structured vehicle trajectories</b> with <b>LLM-generated natural-language annotations</b> for safety-critical driving. K-Risk aggregates <b>20 human-driven (HV) and automated-vehicle (AV) trajectory sources</b> across Europe, China and the United States — covering highways, urban freeways, intersections and roundabouts — and curates <b>31,398 high-risk events</b>, including a <b>1,036-event extreme near-collision subset</b>.The <b>data-processing pipeline</b> (event extraction, scenario-description generators, the closed-loop LLM annotation driver, and the statistics/figure scripts) is distributed separately as source code; see https://github.com/benmagnifico/K-Risk. This archive contains the processed dataset only.<br>

Links and sources

Comments