Improving Out-of-Distribution Generalization of Trajectory Prediction for Autonomous Driving via Polynomial Representations

TL;DR

This work tackles the problem that trajectory prediction models are predominantly evaluated on In-Distribution data and often fail to generalize across datasets. It introduces a dataset homogenization protocol to enable Out-of-Distribution testing between Argoverse 2 and Waymo Open, and proposes Everything Polynomial (EP), a compact predictor that represents inputs with Bernstein polynomials of degree $5$ for histories and maps with degree $3$, while predicting future trajectories as $6$-th degree polynomials using a linear reconstruction with a fixed matrix $H$. EP achieves near-SotA In-Distribution performance with a fraction of the parameters and inference time, and demonstrates significantly improved OoD robustness compared to sequence-based baselines, particularly under homogeneous augmentation and polynomial representations. The results advocate adding OoD testing to standard evaluation and show that careful data homogenization and parametric trajectory representations can markedly enhance cross-dataset generalization for autonomous driving trajectory prediction.

Abstract

Robustness against Out-of-Distribution (OoD) samples is a key performance indicator of a trajectory prediction model. However, the development and ranking of state-of-the-art (SotA) models are driven by their In-Distribution (ID) performance on individual competition datasets. We present an OoD testing protocol that homogenizes datasets and prediction tasks across two large-scale motion datasets. We introduce a novel prediction algorithm based on polynomial representations for agent trajectory and road geometry on both the input and output sides of the model. With a much smaller model size, training effort, and inference time, we reach near SotA performance for ID testing and significantly improve robustness in OoD testing. Within our OoD testing protocol, we further study two augmentation strategies of SotA models and their effects on model generalization. Highlighting the contrast between ID and OoD performance, we suggest adding OoD testing to the evaluation criteria of trajectory prediction models.

Figures (9)

• Figure 1: Kernel density plot of the maximum absolute curvature and length for $5000$ random lane segments in A2 and WO. Contours indicate the $20$-th, $40$-th, $60$-th, and $80$-th percentiles, respectively.
• Figure 2: Our proposed model architecture. Inputs: Agent histories and road geometry are both represented via polynomials. Outputs: The tracked initial states and predicted states are fused into one polynomial trajectory prediction, ensuring continuity of past observation and future prediction.
• Figure 3: The two augmentation strategies for non-focal agent data employed in benchmark models. Left: FMAE employs heterogeneous augmentation, representing information in the focal agent's coordinate frame and only making uni-modal predictions for non-focal agents. Right: QCNet employs homogeneous augmentation, encoding information in each agent's individual coordinate frame and making multi-modal predictions for both focal and non-focal agents alike.
• Figure 4: One traffic scenario in A2 represented with polynomials.
• Figure 5: ID performance of benchmarks on A2. Dashed Lines: Results reported by original authors with competition settings. Bars: Results for retrained benchmark models on homogenized datasets.