Sentence-resampled BERT-CRF model for autonomous vehicle crash causality analysis from large-scale accident narrative...

Area of Science:

• Traffic Safety
• Artificial Intelligence
• Natural Language Processing

Background:

• Autonomous vehicle (AV) safety is paramount, yet crash causality analysis often overlooks unstructured narrative data.
• Existing methods struggle with data imbalance and small sample sizes inherent in crash narratives.

Purpose of the Study:

• To develop a novel framework for analyzing AV crash causality using unstructured crash narratives.
• To identify high-risk crash scenarios and their underlying causal factors.

Main Methods:

• A Sentence-resampled BERT-CRF model was developed to extract causal movement chains (CMC) from narratives.
• A DREAM-inspired hierarchical causal attribution framework was employed for systematic analysis.
• Data imbalance was addressed using sentence-level resampling.

Main Results:

• The model achieved 98.03% accuracy on complete data and 96.14% on 10% of the data.
• Rear-end (48.57%) and lane-change (17.04%) collisions were identified as high-risk scenarios.
• Causal factors include conventional vehicle (CV) proximity, AV decision-making, and intent recognition delays.

Conclusions:

• The proposed framework effectively bridges unstructured data and causal inference for AV safety.
• Key causal factors involve human-computer interaction, environmental perception, and roadway infrastructure.
• Findings support AV manufacturers in optimizing algorithms and authorities in developing regulations.