Uncertainty Matters: Structured Probabilistic Online Mapping for Motion Prediction in Autonomous Driving

Abstract

Online map generation and trajectory prediction are critical components of the autonomous driving perception-prediction-planning pipeline. While modern vectorized mapping models achieve high geometric accuracy, they typically treat map estimation as a deterministic task, discarding structural uncertainty. Existing probabilistic approaches often rely on diagonal covariance matrices, which assume independence between points and fail to capture the strong spatial correlations inherent in road geometry. To address this, we propose a structured probabilistic formulation for online map generation. Our method explicitly models intra-element dependencies by predicting a dense covariance matrix, parameterized via a Low-Rank plus Diagonal (LRPD) covariance decomposition. This formulation represents uncertainty as a combination of a low-rank component, which captures global spatial structure, and a diagonal component representing independent local noise, thereby capturing geometric correlations without the prohibitive computational cost of full covariance matrices. Evaluations on the nuScenes dataset demonstrate that our uncertainty-aware framework yields consistent improvements in online map generation quality compared to deterministic baselines. Furthermore, our approach establishes new state-of-the-art performance for map-based motion prediction, highlighting the critical role of uncertainty in planning tasks. Code is published under link-available-soon.

Figures (3)

• Figure 1: Illustration of different uncertainty representations in probabilistic mapping for a scene with an ego vehicle and a predicted vehicle. (Left) Treating each point within a map element (i.e. a polyline) in an independent manner can result in inconsistent covariance estimates across consecutive points. This stems from the inability to model naturally-present correlations between locations of points, such as a usually higher uncertainty for farther points. A limitation of the fully-diagonal covariance matrix representation present in the literature, this can result in erroneous predictions that in turn impact downstream planning. (Right) A representation that captures correlations between individual points models the underlying uncertainty more accurately. This is enabled by a covariance matrix with off-diagonal elements, which equips the probabilistic model with ability to reason about relationships between points.
• Figure 2: Qualitative comparison of OMG predictions on several nuScenes nuscenes scenarios. We display samples drawn from the predicted distribution as thin transparent lines, while the thick solid line represents the predictive mean. (Left) Baseline models using diagonal covariance (independent uncertainty) fail to capture dependencies between points, resulting in jagged, spatially incoherent samples. (Middle) Our proposed LRPD model explicitly accounts for spatial correlations, producing smooth and geometrically consistent samples. (Right) The Ground Truth HD map for reference.
• Figure 3: Uncertainty calibration of our method. (Left) Visualization of predicted versus ground-truth map element. The prediction aligns closely with the ground truth initially, but the positional error gradually increases toward the end. (Right) Predicted covariance demonstrating that our model's uncertainty correlates with the error. Uncertainty grows correspondingly as the prediction diverges. Specifically, the model correctly outputs low uncertainty for the $y$-coordinate (where the error remains small) and appropriately assigns high uncertainty to the $x$-coordinate as its error increases.