EDRF: Enhanced Driving Risk Field Based on Multimodal Trajectory Prediction and Its Applications

TL;DR

The paper tackles driving risk under uncertainty by introducing the Enhanced Driving Risk Field (EDRF), which fuses multimodal trajectory predictions with Gaussian uncertainty to compute a driving risk probability (DRP) and its consequence via a virtual mass. The DRP is aggregated across predicted trajectories in a Frenet framework to form the EDRF, enabling a unified risk view for traffic risk monitoring, ego-vehicle risk analysis, and motion/trajectory planning. Key contributions include a QCNet-based multimodal prediction backbone, a trajectory-adaptive Gaussian cross-section (and Laplace-like cross-section for ego risk), and a principled definition of interaction risk (IR) used across applications. The approach aims to improve safety and decision-making in autonomous and human-driven driving by explicitly modeling and propagating uncertainty through a field-theoretic risk representation, with plans for large-scale validation on public data and simulators.

Abstract

Driving risk assessment is crucial for both autonomous vehicles and human-driven vehicles. The driving risk can be quantified as the product of the probability that an event (such as collision) will occur and the consequence of that event. However, the probability of events occurring is often difficult to predict due to the uncertainty of drivers' or vehicles' behavior. Traditional methods generally employ kinematic-based approaches to predict the future trajectories of entities, which often yield unrealistic prediction results. In this paper, the Enhanced Driving Risk Field (EDRF) model is proposed, integrating deep learning-based multimodal trajectory prediction results with Gaussian distribution models to quantitatively capture the uncertainty of traffic entities' behavior. The applications of the EDRF are also proposed. It is applied across various tasks (traffic risk monitoring, ego-vehicle risk analysis, and motion and trajectory planning) through the defined concept Interaction Risk (IR). Adequate example scenarios are provided for each application to illustrate the effectiveness of the model.

Figures (12)

• Figure 1: The framework of Enhanced Driving Risk Field ($EDRF$) and its applications.
• Figure 2: Examples of the multimodal trajectory prediction results.The orange line represents the vehicle’s historical trajectory, the red line denotes the ground truth, and the green lines depict the multimodal trajectory prediction results.
• Figure 3: The predicted trajectory and the Frenet coordinate system built upon it.
• Figure 4: The modelling process of the $DRP$. We finally obtain the torus with Gaussian cross-section, whose height decreases and width increases along the predicted trajectory.
• Figure 5: Comparison of deep learning prediction-based and kinematic-based future trajectory prediction methods.