Stay on Track: A Frenet Wrapper to Overcome Off-road Trajectories in Vehicle Motion Prediction

TL;DR

The paper tackles the generalisation gap in vehicle motion prediction where models over-rely on motion history and falter on map-challenging, perturbed scenes. It introduces a model-agnostic Frenet-frame wrapper that represents scene inputs and outputs in lane-centered coordinates for each relevant centreline, querying SotA predictors in parallel across frames and aggregating results. Empirical evaluation on Argoverse with LaneGCN and Multipath++ shows the wrapper dramatically reduces off-road predictions on perturbed scenes (by over 90%) while maintaining or improving trajectory diversity, at a small cost to original-scene accuracy. The approach provides a practical, scalable path to map-aware robustness without altering existing model architectures. Overall, the Frenet wrapper enhances reliability in safety-critical forecasting by aligning predictions with lane topology and diversifying forecasted trajectories according to the scene context.

Abstract

Predicting the future motion of observed vehicles is a crucial enabler for safe autonomous driving. The field of motion prediction has seen large progress recently with state-of-the-Art (sotA) models achieving impressive results on large-scale public benchmarks. However, recent work revealed that learning-based methods are prone to predict off-road trajectories in challenging scenarios. These can be created by perturbing existing scenarios with additional turns in front of the target vehicle while the motion history is left unchanged. We argue that this indicates that SotA models do not consider the map information sufficiently and demonstrate how this can be solved by representing model inputs and outputs in a Frenet frame defined by lane centreline sequences. To this end, we present a general wrapper that leverages a Frenet representation of the scene, and that can be applied to SotA models without changing their architecture. We demonstrate the effectiveness of this approach in a comprehensive benchmark using two SotA motion prediction models. Our experiments show that this reduces the off-road rate on challenging scenarios by more than 90% without sacrificing average performance.

Figures (3)

• Figure 1: Method overview After identifying relevant lane centreline sequences for the TV, the scene is transformed to the corresponding Frenet frame of each centreline. Next, the prediction model is inferred and the output is projected back to the cartesian frame. Finally, predictions from all centrelines are aggregated. For simplicity, the figure shows $K=1$ predictions per centreline.
• Figure 2: scene-attack perturbations. Three different transformations are applied to the original scene. All of them create a turn at a fixed distance ahead of the target vehicle. Observe how the motion history and the ground truth are slowed down to account for the large curvature.
• Figure 3: Qualitative Results From top to bottom two original and two perturbed scenes are shown. Centrelines used by the SD models are shown in green, state history in yellow. Predicted trajectories and ground truth are shown in blue and red, respectively. In the top example, the TV has just passed the branching point so that only the straight centreline is detected. The SD models are still able to predict the right-turn mode, significantly off the reference centreline. The second row demonstrates the increased diversity of SD model predictions. The third row shows deviations from the reference line. In the bottom example, the SD models adapt to a difficult perturbation, while the original models predict only off-road trajectories (Best viewed in colour).