AUTOSIGNAL

Human-curated. Expert-annotated. Every signal traced to source. 人工精选 · 专家点评 · 每条信号可溯源

Research & IP研究与专利

DriveDNA: 4,121-Drive Naturalistic Dataset Reveals Vehicle and Road Shortcuts in Driving-Style RecognitionDriveDNA自然驾驶数据集:4121次行程揭示驾驶风格识别中的车辆/路况捷径问题

S 1.7 T1 1 sources1 个来源 R7-research

Researchers released DriveDNA, a large-scale naturalistic driving benchmark for driving-style identification comprising 4,121 drives from 465 drivers across 115 vehicle models, totaling 975 hours of human-controlled driving with synchronized CAN-bus and forward-video data, evaluated across 30 baseline configurations under one unified framework.
Learned representations reach 93.5% AUROC on unseen-driver re-identification versus 70.7% for classical descriptors, and retain 81.1% accuracy under condition-matched evaluation where the classical descriptors collapse to chance level.
A video-only model matches CAN-based re-identification at 93.7% but predicts the driving route at 347× chance level, and its accuracy collapses to 67.5% once vehicle/road conditions are matched — showing that seemingly strong driver-recognition performance can be a shortcut based on route or vehicle rather than genuine driving style.

研究团队发布DriveDNA,一个用于驾驶风格识别的大规模自然驾驶基准数据集,包含465位驾驶员在115种车型上的4121次行程,总计975小时人工驾驶数据(CAN总线+前向摄像头同步采集),并在统一框架下测试了30种基线配置。
学习到的表征在未见驾驶员重识别任务上达到93.5% AUROC,远超经典描述符的70.7%;在条件匹配评估下仍保持81.1%的准确率,而经典描述符则降至随机水平。
一个仅用视频的模型在重识别任务上达到93.7%,与CAN数据模型持平,但其路线预测准确率高达随机水平的347倍,一旦车辆/路况条件被匹配对齐,准确率骤降至67.5%——说明表面上很强的驾驶员识别能力,可能只是利用了路线或车辆的捷径,而非真实的驾驶风格。

hf-daily-papers