Fast Long-Term Multi-Scenario Prediction for Maneuver Planning at Unsignalized Intersections

TL;DR

This work tackles long-horizon, maneuver-aware motion prediction for unsignalized intersections by decoupling gap-acceptance decisions from velocity control in a dual-stage neural framework. The two small MLPs, MLP_gap and MLP_acc, are trained to predict per-pair gap decisions and longitudinal acceleration, respectively, allowing explicit integration of cooperative maneuver priorities. Trained on real-world inD data with PPO pre-training and GAIL IRL refinement, the method achieves 83.6% gap-acceptance accuracy and 14 m RMSE (MAD 6 m) over a 10 s horizon, while delivering real-time multi-scenario predictions (under 3.5 ms on CPU for 15 vehicles) to support maneuver planning. The approach enables rapid evaluation of numerous potential maneuvers, potentially improving traffic efficiency and safety in infrastructure-supported coordination settings, with future work addressing perception uncertainties and probabilistic scenario predictions.

Abstract

Motion prediction for intelligent vehicles typically focuses on estimating the most probable future evolutions of a traffic scenario. Estimating the gap acceptance, i.e., whether a vehicle merges or crosses before another vehicle with the right of way, is often handled implicitly in the prediction. However, an infrastructure-based maneuver planning can assign artificial priorities between cooperative vehicles, so it needs to evaluate many more potential scenarios. Additionally, the prediction horizon has to be long enough to assess the impact of a maneuver. We, therefore, present a novel long-term prediction approach handling the gap acceptance estimation and the velocity prediction in two separate stages. Thereby, the behavior of regular vehicles as well as priority assignments of cooperative vehicles can be considered. We train both stages on real-world traffic observations to achieve realistic prediction results. Our method has a competitive accuracy and is fast enough to predict a multitude of scenarios in a short time, making it suitable to be used in a maneuver planning framework.

Figures (5)

• Figure 1: Unsignalized intersection near Bendplatz in Aachen, Germany, and a traffic scenario with future vehicle paths, as found in the inD dataset bock_ind_2020. Vehicles with red arrows have to yield. This scenario has 8 pairs of conflicting vehicles and 14 possible outcomes, i.e. vehicle crossing orders.
• Figure 2: Proposed architecture for gap acceptance and motion prediction.
• Figure 3: Illustration of the observation input features listed in Table \ref{['tbl:agent_observation']}.
• Figure 4: Prediction accuracy of our $\text{MLP}_\text{acc}$ vs. IDM on the validation dataset. For both, our gap acceptance model was used. The root mean square error (RMSE) and median absolute deviation (MAD) per vehicle is shown.
• Figure 5: Runtimes for a 10s prediction horizon at typical numbers of vehicles in the intersection scenario, averaged over 2000 random scenes.