Monocular Localization With Vector HD Maps Using Geometric-Context Data Association

Abstract

Due to their low cost and wide field of view, monocular cameras hold great promise for visual localization. Nevertheless, significant challenges persist in associating detected semantic features with map landmarks, arising from sensing noise, depth cues, and dynamic driving scenes. Most existing approaches neglect the contextual properties of landmarks, which may result in incorrect data association and degrade localization performance. In response, this paper proposes a monocular localization method with vector High-Definition (HD) maps using geometric-context for data association. The method integrates local positions, spatial layout, and geometric shapes of landmarks to construct a normalized weighted error model, while incorporating an outlier removal strategy and applying the Hungarian algorithm for robust association. In the optimization phase, ground plane information is extracted from the vector HD maps, and a structural score-based ground plane constraint is introduced to improve z-axis estimation accuracy. Based on the proposed approach, we build a flexible and modular localization system to accommodate various sensor configurations and onboard computational resources. We conducted extensive experiments in typical urban road scenarios as well as challenging spiral on-ramp scenarios, and the results demonstrate the system's accuracy and robustness in autonomous urban localization.