AMFormer-based framework for accident responsibility attribution: Interpretable analysis with traffic accident...

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Traffic accident responsibility determination traditionally relies on subjective evidence, leading to potential biases and inaccuracies.
• Existing methods lack objectivity and struggle with the complexity of accident scenarios.

Purpose of the Study:

• To introduce AMFormer (Arithmetic Feature Interaction Transformer), a novel deep learning framework for accurate and interpretable traffic accident responsibility prediction.
• To enhance objectivity and reduce bias in assigning responsibility for traffic incidents.

Main Methods:

• Developed AMFormer, a deep learning model utilizing spatiotemporal feature modeling to capture intricate factor interactions.
• Applied SHAP (SHapley Additive Interpretation) analysis for feature attribution and transparency.
• Validated the model on real-world traffic accident datasets.

Main Results:

• AMFormer achieved a 93.46% accuracy and a 93% F1-Score in predicting accident responsibility.
• The model demonstrated superior performance compared to traditional methods and other deep learning approaches.
• SHAP analysis identified key features influencing responsibility attribution and their combined effects.

Conclusions:

• AMFormer offers a robust and interpretable solution for traffic accident responsibility prediction.
• The framework enhances the credibility of responsibility attribution and has implications for autonomous vehicle safety.
• This research provides a foundation for more objective and data-driven approaches in traffic safety analysis.