Online Navigation Refinement: Achieving Lane-Level Guidance by Associating Standard-Definition and Online Perception Maps
Jiaxu Wan, Xu Wang, Mengwei Xie, Xinyuan Chang, Xinran Liu, Zheng Pan, Mu Xu, Hong Zhang, Ding Yuan, Yifan Yang
TL;DR
This paper introduces Online Navigation Refinement (ONR), a method that refines standard maps into precise lane-level guidance by associating them with real-time perception maps, overcoming traditional map limitations.
Contribution
It presents a novel map association dataset, a transformer-based alignment model, and an evaluation metric, enabling accurate lane-level navigation without extensive HD maps.
Findings
MAT achieves 34 ms latency, outperforming existing methods.
The dataset contains 30K scenarios with 2.6M lane annotations.
ONR enables low-cost, real-time lane-level guidance.
Abstract
Lane-level navigation is critical for geographic information systems and navigation-based tasks, offering finer-grained guidance than road-level navigation by standard definition (SD) maps. However, it currently relies on expansive global HD maps that cannot adapt to dynamic road conditions. Recently, online perception (OP) maps have become research hotspots, providing real-time geometry as an alternative, but lack the global topology needed for navigation. To address these issues, Online Navigation Refinement (ONR), a new mission is introduced that refines SD-map-based road-level routes into accurate lane-level navigation by associating SD maps with OP maps. The map-to-map association to handle many-to-one lane-to-road mappings under two key challenges: (1) no public dataset provides lane-to-road correspondences; (2) severe misalignment from spatial fluctuations, semantic disparities,…
Peer Reviews
Decision·ICLR 2026 Poster
- Although existing autonomous driving datasets, such as nuScenes and OpenLane-V2, provide fine-grained local lane geometry annotations, and OpenStreetMap provides road-level SD Map annotations, the mappings between them are less explored. The OMA dataset proposed by this work greatly mitigates this data gap. - Path-aware attention is introduced to align the topologies under the distractions introduced by spatial fluctuations and semantic disparities. By forcing each token can appear only once i
- Figure 5 is not self-contained. The concepts of path-aware and spatial attention are hard to grasp from the visual alone, as the figure lacks sufficient illustrative descriptions or a detailed caption. - Abbreviations are sometimes unclear or non-standard. For example, "Para." in Table 1 is ambiguous and should be explicitly defined (e.g., as "Parameters"). - The section on "attention with vector serialization" is underdeveloped. The explanation is too brief, lacking the detail needed for the
1. It directly tackles the core pain points of existing lane-level navigation solutions. The proposed ONR task bridges SD and OP maps to achieve low-cost, up-to-date lane-level guidance, which is highly relevant to real-world needs in GIS and autonomous driving. 2. The paper provides a "dataset-model-metric" trinity solution to fill research gaps:OMA Dataset, MAT Model and NR P-R Metric. 3. The paper provided solid experimental validation.
1. The test set's OP map noise only comes from MapTRv2. It does not evaluate MAT’s performance under other common OP noise types (e.g., severe lane occlusion by vehicles, sensor failure in heavy rain/fog), I think it should be evaluated to improve the model's robustness. 2. The test set uses OP maps generated solely by MapTRv2. If other OP map generators produce different noise patterns, whether MAT’s performance will degrade or not remains untested, i think the cross-generator generalization a
1. This paper proposes a new and important task for the online perception maps association, which is valuable for real-world applications in autonomous driving. Assigning prediction lanes to SD road elements is essential for building topological relationships between lanes and roads, further benefiting navigation and planning. Building a corresponding benchmark and evaluation metric fills the gap in the field of online map construction, as far as I know. 2. The evaluation metrics are carefully d
1. For the subfigure Fig 2.(3), there are two junction curve lines connecting two crossroads. The mapping of such transition lines is ambiguous due to their connection attributes. How to solve this problem when constructing the ground truth assignment labels? 2. As this paper also studies the noise of road elements and assignment with predicted lanes, a line of related works regarding the use of SDMaps and the noise problem for map perception should be included and discussed, such as SMERF [1],
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAutomated Road and Building Extraction · Autonomous Vehicle Technology and Safety · Data Management and Algorithms