Abstract
Abstract
High-definition (HD) vector maps are essential to trajectory forecasting, yet existing methods often use them only as passive context and remain dependent on identity-consistent tracking histories. We propose a unified geometry-centric forecasting framework that treats road-space geometry as an active computational substrate for both future prediction and history recovery under tracking-available and no-tracking conditions. The framework uses a common road-space representation and topology-guided decoupled decoder architecture across both settings, while differing only in how motion history is constructed. With reliable tracking, agent history is incrementally aggregated across shifting local frames for efficient streaming inference. Without tracking, frame-wise detections are projected into lane-space occupancy memories, from which target-specific history is recovered through geometry-conditioned querying. The decoder then generates topology-consistent motion intents and realizes them as final trajectories using surrounding-agent interactions. Experiments on Argoverse 2 (AV2), Argoverse 1 (AV1), and the Waymo Open Motion Dataset (WOMD) show that the proposed framework achieves competitive performance under reliable tracking and supports low-latency streaming inference, reaching 44.32 ms per frame with pre-encoded maps. Under no-tracking conditions, the occupancy-based pathway preserves most of the tracking-available performance on the primary K = 6 metrics, with relative increases in minFDE 6 , minADE 6 , and MR 6 limited to 6%–8% on AV2 and below 3% on AV1 and WOMD. Additional validation on real sensor data with annotated detections and a tracking-by-detection association pipeline shows that the occupancy-based pathway remains robust to tracking identity errors, whereas identity-dependent methods degrade substantially. These results demonstrate that actively exploiting road-space geometry can substantially reduce trajectory forecasting’s reliance on upstream tracking.
Direct answer
What can I do from this paper page?
Use this page to scan "Road space as memory and inference: A unified geometry-centric framework for trajectory forecasting" quickly: start with the summary and abstract, then check the authors, source, topics, and related papers. From here, open Scollr to follow Autonomous Vehicle Technology and Safety research, save the paper, or map adjacent work.
Research areas
Follow related topics
Citation
BibTeX
@article{Xu2026Road,
title = {Road space as memory and inference: A unified geometry-centric framework for trajectory forecasting},
author = {Yiming Xu and Hao Cheng and Monika Sester},
journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
year = {2026},
doi = {10.1016/j.isprsjprs.2026.10.023},
url = {https://doi.org/10.1016/j.isprsjprs.2026.10.023}
}
FAQ
Using this paper in a discovery workflow
How do I find related work for this paper?
Use the related papers and topic links on this page as starting points. In Scollr, you can also open the paper and build a literature map around its references, citing papers, and related work.
How can I keep up with new Autonomous Vehicle Technology and Safety research papers?
Follow Autonomous Vehicle Technology and Safety research in Scollr. New papers from the topic flow into a personalized feed, and you can save useful studies to revisit later.
Can I cite this paper from this page?
This page includes a static BibTeX block for Road space as memory and inference: A unified geometry-centric framework for trajectory forecasting. Always verify the DOI, source, and publication details against the publisher record before submitting a manuscript.
Follow this research in Scollr
Follow the topics and authors behind this paper, save useful studies, and build a literature map when you are ready to go deeper.
Get the app