Uncertainty-aware spatiotemporal interaction learning for pre-conflict risk evolution with a risk-increase prior

Area of Science:

• Engineering
• Computer Science
• Transportation Safety

Background:

• Current methods for dynamic risk assessment in vehicles rely on static views and incomplete uncertainty modeling.
• This limits the ability to accurately track the evolution of conflict risk over time.
• Existing metrics like Time To Collision (TTC) have limitations in early and reliable risk detection.

Purpose of the Study:

• To develop a new risk quantification model that integrates driver control inputs and multi-vehicle spatiotemporal interactions.
• To explicitly model uncertainty in risk estimation to better understand conflict evolution.
• To establish a proactive safety assessment paradigm for vehicles.

Main Methods:

• A novel risk quantification model was developed, incorporating driver control inputs and multi-vehicle spatiotemporal data.
• The model explicitly outputs uncertainty estimates alongside risk predictions.
• Performance was evaluated against existing metrics like TTC, DRAC, PSD, ACT, and EI across various driving scenarios.

Main Results:

• The proposed model demonstrated superior risk discrimination compared to TTC, DRAC, PSD, ACT, and EI.
• It detected elevated conflict risk an average of 1.15 seconds before the conflict point.
• In representative scenarios, the model showed a lower false alarm rate than TTC and perceived rising risk 1.44 seconds earlier.

Conclusions:

• The developed model provides a robust framework for real-time conflict risk quantification and evolution tracking.
• Explicit uncertainty outputs enable reliable capture of risk dynamics and support model calibration.
• The findings establish a new paradigm for proactive vehicle safety assessment by jointly estimating risk and confidence.