LaneDiffusion: Improving Centerline Graph Learning via Prior Injected BEV Feature Generation

2025-11-09

AI Summary

Traditional centerline graph learning methods predominantly employ deterministic modeling, exhibiting limited spatial reasoning capability and poor robustness to occlusions and invisible lanes. To address these limitations, this work introduces diffusion models to centerline graph learning for the first time. Specifically, we design a Lane Prior Injection (LPI) module and a Prior-Driven Diffusion (LPD) module operating on bird’s-eye-view (BEV) features, enabling structural-aware generative modeling. Our approach leverages prior-guided iterative denoising to significantly enhance topological recovery and spatial reasoning under occlusion. Extensive experiments on nuScenes and Argoverse2 demonstrate state-of-the-art performance across both point-level and segment-level metrics. These results validate the effectiveness and robustness of the generative paradigm for centerline graph learning.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Graph-based Machine Learning

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Centerline graphs, crucial for path planning in autonomous driving, are traditionally learned using deterministic methods. However, these methods often lack spatial reasoning and struggle with occluded or invisible centerlines. Generative approaches, despite their potential, remain underexplored in this domain. We introduce LaneDiffusion, a novel generative paradigm for centerline graph learning. LaneDiffusion innovatively employs diffusion models to generate lane centerline priors at the Bird's Eye View (BEV) feature level, instead of directly predicting vectorized centerlines. Our method integrates a Lane Prior Injection Module (LPIM) and a Lane Prior Diffusion Module (LPDM) to effectively construct diffusion targets and manage the diffusion process. Furthermore, vectorized centerlines and topologies are then decoded from these prior-injected BEV features. Extensive evaluations on the nuScenes and Argoverse2 datasets demonstrate that LaneDiffusion significantly outperforms existing methods, achieving improvements of 4.2%, 4.6%, 4.7%, 6.4% and 1.8% on fine-grained point-level metrics (GEO F1, TOPO F1, JTOPO F1, APLS and SDA) and 2.3%, 6.4%, 6.8% and 2.1% on segment-level metrics (IoU, mAP_cf, DET_l and TOP_ll). These results establish state-of-the-art performance in centerline graph learning, offering new insights into generative models for this task.

Problem

• Generating centerline graphs for autonomous driving path planning
• Overcoming limitations of deterministic methods with spatial reasoning
• Addressing occluded centerline prediction using diffusion models

Innovation

• Uses diffusion models to generate BEV lane priors
• Integrates Lane Prior Injection and Diffusion Modules
• Decodes vectorized centerlines from prior-injected features