Producing and Leveraging Online Map Uncertainty in Trajectory Prediction

TL;DR

The work addresses the lack of uncertainty in online vectorized HD maps used for autonomous driving. It introduces a Laplace-based uncertainty head for map vertices, extends multiple online map estimators to output ($oldsymbol{bc}, oldsymbol{b}$) alongside coordinates, and trains with a Negative Log-Likelihood loss, preserving mapping performance. Downstream trajectory prediction models (DenseTNT, HiVT) are augmented to ingest map uncertainty, improving endpoint accuracy and reducing miss rates, while also accelerating training convergence by up to ~50%. Experiments on nuScenes demonstrate robust benefits across 16 mapping/prediction combinations, with occlusion, distance, and weather effects captured by the uncertainty estimates. The findings highlight the practical benefit of uncertainty-aware online mapping for more reliable, faster-to-train autonomous driving systems and point to future direction in calibrating map-model uncertainty.

Abstract

High-definition (HD) maps have played an integral role in the development of modern autonomous vehicle (AV) stacks, albeit with high associated labeling and maintenance costs. As a result, many recent works have proposed methods for estimating HD maps online from sensor data, enabling AVs to operate outside of previously-mapped regions. However, current online map estimation approaches are developed in isolation of their downstream tasks, complicating their integration in AV stacks. In particular, they do not produce uncertainty or confidence estimates. In this work, we extend multiple state-of-the-art online map estimation methods to additionally estimate uncertainty and show how this enables more tightly integrating online mapping with trajectory forecasting. In doing so, we find that incorporating uncertainty yields up to 50% faster training convergence and up to 15% better prediction performance on the real-world nuScenes driving dataset.

Figures (14)

• Figure 1: Producing uncertainty from online HD map estimation methods and incorporating it in downstream modules yields a variety of benefits. Left: Ground truth HD map and agent positions. Middle: HiVT zhou2022hivt predictions using the map output by MapTR MapTR. Right: HiVT zhou2022hivt predictions using the map output by MapTR MapTR augmented with point uncertainties (which are large as the left road boundary is occluded by parked vehicles).
• Figure 2: Many online HD vector map estimation methods operate by encoding multi-camera images, transforming them to a common BEV feature space, and regressing map element vertices. Our work augments this common output structure with a probabilistic regression head, modeling each map vertex as a Laplace distribution. To assess the resulting downstream effects, we further extend downstream prediction models to encode map uncertainty, augmenting both GNN-based and Transformer-based map encoders.
• Figure 3: Our proposed uncertainty formulation is able to capture uncertainty stemming from occlusions between the AV's cameras and surrounding map elements. Left: Images from the front and front-right cameras. Right: HD maps from our augmented online HD mapping models. Ellipses show the std. dev. of distributions. Colors are road boundary, lane divider, pedestrian crossing, lane centerline.
• Figure 4: In a dense parking lot, many models fail to produce accurate maps. Left: Images from the rear and rear-left cameras. Right: HD maps from our augmented online HD mapping models. Ellipses show the std. dev. of distributions. Colors indicate road boundary, lane divider, pedestrian crossing, lane centerline.
• Figure 5: Our uncertainty formulation captures the fact that uncertainty generally increases with the distance between the predicted map elements and the AV, owing to the difficulty of resolving the details of faraway objects in images. Error bars show 95% confidence intervals.