Publication Date

Spring 2026

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering

Advisor

Kaikai Liu; Bernardo Flores; Jun Liu

Abstract

Accurate online estimation of road structures is essential for safe autonomous driving. Existing systems rely either on costly high-definition (HD) maps or purely sensor-based online mapping, both with inherent limitations. Prior-informed approaches integrate external map priors with onboard perception, but full HD map priors suffer from shortcut learning, where the model memorizes prior geometry rather than learning to infer road structure, causing severe performance degradation when priors are unavailable at deployment. This thesis presents two complementary contributions. First, I propose replacing full HD map priors with a lightweight, non-visible driveline prior encoding only lane-centric topology. Because drivelines are structurally adjacent to but distinct from prediction targets, the network must reason from topology to geometry rather than copy prior elements, yielding improved generalization: models trained with driveline priors retain performance when priors are absent at inference (−0.9 mAP drop), compared to the baseline UPPM HD-map prior model (−7.6 mAP drop) on nuScenes. Second, I integrate a diffusion-based denoising model to refine the prior-augmented bird's-eye-view (BEV) feature representations. Adapted from BEVDiffuser, the model is re-conditioned on vectorized polyline layouts rather than 3D object bounding boxes, enabling layout-guided denoising with no additional inference cost. Experiments on nuScenes demonstrate that the integrated system achieves 78.2 mAP (+4.2 over the driveline-only baseline), surpassing the HD-map prior model (76.1 mAP) without requiring any HD map information.

Download

Available for download on Sunday, July 25, 2027

Share

COinS