Quantifying Uncertainty in Motion Prediction with Variational Bayesian Mixture

TL;DR

SeNeVA introduces a variational Bayesian mixture model for uncertainty-aware single-agent motion prediction, conditioning on history and scene context. It decomposes futures into K modality-specific components with a shared encoder, latent temporal dynamics, and an assignment network to estimate mixture weights, enabling efficient sampling via NMS. Empirical results on INTERACTION and Argoverse 2 show competitive prediction accuracy and well-calibrated uncertainty, with improved OOD detection through entropy and faster inference compared to deep ensembles. The approach uses a compact model (~1.3M parameters) and provides representative trajectory samples suitable for safety-critical evaluation and downstream decision-making.

Abstract

Safety and robustness are crucial factors in developing trustworthy autonomous vehicles. One essential aspect of addressing these factors is to equip vehicles with the capability to predict future trajectories for all moving objects in the surroundings and quantify prediction uncertainties. In this paper, we propose the Sequential Neural Variational Agent (SeNeVA), a generative model that describes the distribution of future trajectories for a single moving object. Our approach can distinguish Out-of-Distribution data while quantifying uncertainty and achieving competitive performance compared to state-of-the-art methods on the Argoverse 2 and INTERACTION datasets. Specifically, a 0.446 meters minimum Final Displacement Error, a 0.203 meters minimum Average Displacement Error, and a 5.35% Miss Rate are achieved on the INTERACTION test set. Extensive qualitative and quantitative analysis is also provided to evaluate the proposed model. Our open-source code is available at https://github.com/PurdueDigitalTwin/seneva.

Figures (8)

• Figure 1: Illustration of multi-modal trajectory distribution. There are four surrounding vehicles (blue) and one target vehicle (red) in the scene. The target vehicle can either turn left within the roundabout (green) or turn right into the nearest exit (orange).
• Figure 2: Architecture of the proposed SeNeVA model. The track and map encoders encode HD map and agent history trajectories. A global encoder module with a cascade of multi-head attention layers passes messages between map and agents to compute the context feature $\boldsymbol{x}$ from the perspective of the target agent. A variational Bayes Model of $K$ components estimates the distribution $p(\boldsymbol{s}_f|\boldsymbol{x})$ of trajectories conditioned on the context feature $x$. Additionally, we have an assignment network to estimate the distribution of mixture coefficients $p(z|\boldsymbol{x})$ conditioned on the context feature. The estimated distributions quantify the uncertainty of all possible future trajectories and enable the sampling of representative ones.
• Figure 3: The graphical representation of the (a) generative model and the (b) variational family. Shaded and unshaded nodes are the observed and latent random variables. Diamond nodes are the model parameters.
• Figure 4: Predicted uncertainty in different road geometry and data distributions. The predicted uncertainty in OOD cases is generally higher than in in-distribution cases.
• Figure 5: Example visualization of quantified uncertainty. We visualize the predicted uncertainty in an in-distribution test case (left) and an OOD test case (right). The heatmap reflects the log-likelihood of a location on the map being visited in the future.