Abstract
Predicting driver intent is a critical challenge for autonomous vehicles (AVs) during complex maneuvers like highway merging. This paper introduces a Dynamic Bayesian Network (DBN) framework that leverages collaborative data from connected vehicles to estimate merge intentions with high precision. By focusing on situational context—specifically relative distance and acceleration—the model achieves over 95% accuracy. Validated using the NGSIM dataset, this research provides a scalable foundation for centralized traffic management systems in autonomous environments.
Introduction: The Challenge of Collaborative Merging
Merging onto a high-speed highway from an on-ramp is a cognitively demanding task requiring precise coordination and gap acceptance. While individual AVs rely on on-board sensors, these often suffer from limited range and line-of-sight obstructions. Collaborative autonomy addresses this by using a centralized system to process data shared between vehicles, facilitating safer and smoother transitions.
Methodology
The proposed system utilizes a Dynamic Bayesian Network (DBN) structured into three layers to estimate driver intentions:
- Vector X (Vehicle Data): Captures raw physical states such as position, speed, and acceleration.
- Vector C (Situational Context): Transforms raw data into relational features, including distance to the merge point and relative metrics compared to surrounding vehicles.
- Output I (Intention): Predicts the probability of a vehicle merging or maintaining priority.
A Nominal Logistic Regression Model (LRM) is employed to map the situational context to a probability value between 0 and 1. The model was trained using Maximum Likelihood Estimation (MLE) on the Next Generation Simulation (NGSIM) dataset from Interstate 80 in California.
Key Insights: Context Over Speed
A significant finding of this research is that absolute speed is a poor predictor of merge intent. Instead, drivers primarily base their decisions on the "gap"—the relative distance and acceleration compared to the vehicle ahead in the main lane. By prioritizing these contextual states, the DBN reduces computational overhead while maintaining high fidelity to human behavior.
Results and Validation
The model demonstrated exceptional performance in predicting real-world driver behavior. It achieved an average accuracy exceeding 95%, with a Root-Mean-Square Error (RMSE) of 0.2168 for the main lane and 0.1674 for the merge lane.
As shown in the performance plots, the LRM output closely tracks the actual driver decisions over time. The clear statistical margin (averaging 0.615) between the probabilities of vehicles in different lanes allows a centralized system to make reliable "go/no-go" decisions for autonomous coordination.
Conclusion and Future Work
The DBN framework offers a semantically flexible approach to intent prediction that can be adapted to various road topologies, such as urban intersections. While highly accurate, future developments must address communication reliability, such as latency in Maneuver Coordination Messages (MCM), and account for the occasional stochastic nature of human driving behavior.
Summary updated