Researchers at Tsinghua University developed PriorFusion, a unified framework that integrates semantic, geometric, and generative shape priors to significantly improve the accuracy and stability of road element perception in autonomous driving systems. The research addresses a long-standing challenge: existing end-to-end perception models often generate irregular shapes, fragmented boundaries, and incomplete road elements in complex urban scenarios.
The team published their study in Communications in Transportation Research on November 18, 2025.
"We design PriorFusion to introduce shape priors into every stage of the perception pipeline, ensuring that road elements remain stable, smooth, and structurally consistent—even under challenging conditions," says Xuewei Tang, first author and researcher at the School of Vehicle and Mobility at Tsinghua University.
Three technical building blocks turn common-sense into reliable road geometry. First, a Shape-Prior-Guided Query Refinement module aligns segmentation masks with vector instances, preventing the mask and polyline from "drifting apart". Second, a low-dimensional shape template space clusters tens of thousands of annotated boundaries into a compact set of anchor shapes that serve as geometric "dictionaries". A lightweight diffusion decoder uses these anchors as initial constraints and iteratively infers missing parts when objects occlude the view.
Performance gains in highly complex environments
In extensive experiments on the widely used nuScenes dataset, the researchers found that PriorFusion significantly outperforms state-of-the-art methods—particularly in difficult scenarios such as intersections, merges, and occluded areas.
The method achieved an overall mAP of 70.4%, surpassing previous works by a wide margin.
Under a stricter threshold of τ = 0.2, which evaluates fine-grained shape fidelity, PriorFusion-V2 improves accuracy by over 7%, demonstrating strong resilience in high-precision conditions.
"Our visualization results show cleaner boundaries, more coherent curves, and far fewer missed detections. PriorFusion maintains structural regularity even when input observations are incomplete," says Mengmeng Yang, co-author and assistant research professor at Tsinghua University.
Shape templates reveal typical patterns
In the study, the researchers constructed a shape template space using tens of thousands of annotated road elements.
This eigenspace captures the typical geometric patterns of lane dividers, road boundaries, and pedestrian crossings—such as smooth curvatures for boundaries and parallel stripe patterns for crossings.
By clustering these low-dimensional descriptors, PriorFusion generates anchor shapes that reflect the most frequent road structures in urban environments. These anchors act as strong geometric priors for the vector decoder, improving convergence and stability.
A lightweight generative enhancement for real-time perception
Unlike conventional diffusion models—which require dozens or hundreds of denoising steps—PriorFusion uses a truncated diffusion process with only two steps.
This design introduces generative priors while still achieving real-time performance on GPU hardware.
"The truncated diffusion strategy keeps computation practical for onboard systems while still offering the benefits of generative modeling. It effectively regularizes predictions when road geometry is partially occluded or ambiguous," explains Xuewei Tang.
Impact and implications
The authors emphasize that PriorFusion is plug-and-play and can be integrated into many existing BEV-based perception frameworks.
With cities increasingly pursuing autonomous driving deployments, reliable road element extraction is vital for downstream planning, HD-map updates, and safe decision-making.
The framework's ability to use priors learned from data—rather than relying solely on outdated HD maps—offers a promising direction for next-generation, map-light autonomous driving systems.
The above research is published in Communications in Transportation Research (COMMTR), which is a fully open access journal co-published by Tsinghua University Press and Elsevier. COMMTR publishes peer-reviewed high-quality research representing important advances of significance to emerging transport systems. COMMTR is also among the first transportation journals to make the Replication Package mandatory to facilitate researchers, practitioners, and the general public in understanding and advancing existing knowledge. At its discretion, Tsinghua University Press will pay the open access fee for all published papers in 2025.
About Communications in Transportation Research
Communications in Transportation Research was launched in 2021, with academic support provided by Tsinghua University and China Intelligent Transportation Systems Association. The Editors-in-Chief are Professor Xiaobo Qu, a member of the Academia Europaea from Tsinghua University and Professor Shuai'an Wang from Hong Kong Polytechnic University. The journal mainly publishes high-quality, original research and review articles that are of significant importance to emerging transportation systems, aiming to serve as an international platform for showcasing and exchanging innovative achievements in transportation and related fields, fostering academic exchange and development between China and the global community.
It has been indexed in SCIE, SSCI, Ei Compendex, Scopus, CSTPCD, CSCD, OAJ, DOAJ, TRID and other databases. It was selected as Q1 Top Journal in the Engineering and Technology category of the Chinese Academy of Sciences (CAS) Journal Ranking List. In 2022, it was selected as a High-Starting-Point new journal project of the "China Science and Technology Journal Excellence Action Plan". In 2024, it was selected as the Support the Development Project of "High-Level International Scientific and Technological Journals". The same year, it was also chosen as an English Journal Tier Project of the "China Science and Technology Journal Excellence Action Plan Phase Ⅱ". In 2024, it received the first impact factor (2023 IF) of 12.5, ranking Top1 (1/58, Q1) among all journals in "TRANSPORTATION" category. In 2025, its 2024 IF was announced as 14.5, maintaining the Top 1 position (1/61, Q1) in the same category. Tsinghua University Press will cover the open access fee for all published papers in 2025.
