HisTrackMap: Global Vectorized High-Definition Map Construction via History Map Tracking
Jing Yang, Sen Yang, Xiao Tan, Hanli Wang · 2025
Read HisTrackMap: Global Vectorized High-Definition Map Construction via History Map Tracking by Jing Yang, Sen Yang, Xiao Tan, Hanli Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-07168.txt.
Introduction
Jing Yang and co-authors present HisTrackMap, a method that builds high detail road maps for self driving cars by tracking map parts over time, in a 2025 paper. Query based builders can waver from frame to frame, which shakes trust in the map. Their frame stores past sightings in a fine raster history.
A fusion step feeds that history back as a prior for each tracked part, smoothing the map as the car moves. The team also coin a global score for time based map quality, filling a gap in current grading. On two large driving data sets the method beat prior bests on both single frame and time based scores.
Catalog record
| CREATOR | Jing Yang, Sen Yang, Xiao Tan, Hanli Wang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-07168
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
HisTrackMap: Global Vectorized High-Definition Map Construction via History Map Tracking Jing Yang, Sen Yang, Xiao Tan, Hanli Wang Abstract As an essential component of autonomous driving systems, high-definition (HD) maps provide rich and precise environmental information for auto-driving scenarios; however, existing methods, which primarily rely on query-based detection frameworks to directly model map elements or implicitly propagate queries over time, often struggle to maintain consistent temporal perception outcomes. These inconsistencies pose significant challenges to the stability and reliability of real-world autonomous driving and map data collection systems. To address this limitation, we propose a novel end-to-end tracking framework for global map construction by temporally tracking map elements’ historical trajectories. Firstly, instance-level historical rasterization map representation is designed to explicitly store previous perception results, which can control and maintain different global instances’ history information in a fine-grained way. Secondly, we introduce a Map-Trajectory Prior Fusion module within this tracking framework, leveraging historical priors for tracked instances to improve temporal smoothness and continuity. Thirdly, we propose a global perspective metric to evaluate the quality of temporal geometry construction in HD maps, filling the gap in current metrics for assessing global geometric perception results. Substantial experiments on the nuScenes and Argoverse2 datasets demonstrate that the proposed method outperforms state-of-the-art (SOTA) methods in both single-frame and temporal metrics. The project page is available at: https://yj772881654.github.io/HisTrackMap. 1 Introduction High-definition (HD) maps, which include vectorized map elements such as lane dividers, pedestrian crossings, and road boundaries, play a critical role in the navigation and planning of autonomous driving [ 32 , 31 , 33 , 23 , 11 ] . Traditional map construction methods use the SLAM-based method [ 25 , 26 , 39 ] to collect offline map data, followed by extensive post-processing to generate HD maps. However, these methods are constrained by significant limitations, including substantial costs, the absence of real-time processing capabilities, and difficulties in accommodating dynamic environments and road updates. Recent advancements in Perspective View (PV)-to-Bird’s-Eye View (BEV) methods have significantly enhanced vectorized HD map construction such as [ 10 , 14 , 13 , 7 ] . These approaches leverage the DETR-based detection paradigm [ 2 ] to achieve precise HD map generation. To illustrate the differences across various paradigms. Nevertheless, internal prediction instabilities within the model coupled with uncontrollable environmental factors, such as occlusions or low-light conditions, frequently result in temporal perception inconsistencies, posing substantial challenges for real-world autonomous driving scenarios. Latent query embedding is used as a stream memory [ 29 , 37 ] , facilitating the propagation of temporal information within a unified latent memory. Furthermore, MapTracker [ 3 ] employs a tracking paradigm that further utilizes query propagation and implicit latent memory to associate instance-level map elements across consecutive frames, enhancing temporal consistency. In such a framework, a Motion MLP proposed in [ 37 ] is essential for obtaining the next-frame latent memory considering the pose transformation between consecutive frames. Consequently, it employs a transformation loss to guide the Motion MLP, implicitly ensuring the accuracy of the pose transformation. It is understood that a substantial amount of historical perception map data generated during vehicle navigation could potentially be leveraged as prior knowledge or utilized to reduce computational overhead in overlapping map regions. However, the current approach of implicit encoding through instance queries remains inadequate for precisely recording or preserving past geometric map information. To fully utilize existing perception results, we propose HisTrackMap, which explicitly maintains instance-level history maps for corresponding map instances under a tracking paradigm. This approach provides prior information for subsequent navigation perception, ensuring smoother and more continuous HD map construction and enhancing the efficiency and accuracy of temporal information propagation. Figure 1 : The global rasterized map with history maps. Red and green represent divider and boundary categories, respectively. In the history maps, different colors indicate unique tracked map instances. The pink instance is tracked at \mathbf{T}_{1} , consistently propagating temporal geometric information and providing priors for the corresponding track query at \mathbf{T}_{2} . Fig. 1 illustrates the global rasterized map generated from a continuous navigation sequence, along with details of instance tracking. Furthermore, we propose Map-Trajectory Prior Fusion, which leverages the history map to provide refined prior information for track queries at the subsequent timestamp by integrating position-aligned PV and BEV features. Our HisTrackMap leverages instance-level history maps to establish one-to-one correspondences between track queries and map trajectories, thereby enhancing global geometric consistency and constructing a temporally consistent vectorized HD map. Although traditional Chamfer Distance mean Average Precision (mAP) metrics are widely used for HD map construction, we argue that single-frame metrics are insufficient for evaluating performance in practical autopilot scenarios and map data collection processes. For example, they fail to maintain global constructed map consistency across sequential frames, which is critical to ensure robust long-term perception, reliable decision-making, and efficient data acquisition. While MapTracker introduced consistency-aware metrics (C-mAP) to penalize inconsistent map elements, they remain indirect methods of evaluation. To address this issue, we propose a global geometric-aware metric (G-mAP), which directly evaluates the global perceptual quality of the entire scene. The proposed HisTrackMap consistently outperforms existing methods, demonstrating the effectiveness of utilizing history maps compared to implicit transformations. In summary, our main contributions are as follows: • An end-to-end tracking-enhanced framework, HisTrackMap, for vectorized HD map construction is proposed, which leverages temporal information by maintaining instance-level history maps, reducing redundant computations, and enhancing efficiency. • The Map-Trajectory Prior Fusion module is designed to fully utilize the map trajectory information from the history map, combined with instance-level perception features, to provide priors for the corresponding track queries in future frames, thereby optimizing the temporal propagation process. • A novel benchmark with a global geometric perspective is introduced to address the limitations of existing evaluation methods in practical applications. Our HisTrackMap achieves SOTA results in the popular nuScenes [ 1 ] and Argoverse 2 [ 30 ] datasets. 2 Related Work Map Perception with Single Frame. Recently, with advances in PV-to-BEV methods [ 34 , 12 , 22 , 19 , 20 , 18 ] , the HD map construction has predominantly been formulated as a detection task using surround-view images from vehicle-mounted cameras in a single-frame setting. MapTR [ 13 , 14 ] proposes a unified approach for map elements, addressing ambiguities and ensuring stable learning. In representing map elements, PivotNet [ 5 ] adopts a pivot-based representation, while GeMap [ 41 ] models geometric features through Euclidean shapes. MGMapNet [ 35 ] leverages multi-granularity representations to model map instances, integrating instance-level and point-level queries. Map Perception with Priors. Some of the latest methods have also focused on integrating rasterized results and standard (SD) maps to provide priors for HD map construction. SMERF [ 21 ] integrates SD maps for online map prediction, with PMapNet [ 9 ] enhancing performance by focusing on relevant SD Map skeletons. TopoSD [ 36 ] encodes SD map elements into neural representations and instance tokens, using them as priors for lane geometry and topology decoding. Mask2Map [ 4 ] uses a two-stage framework, generating a rasterized map via a segmentation network and refining it into a vectorized map with a mask-driven network. HRMapNet [ 40 ] leverages historical rasterized maps at the city scale and vehicle