TopoStreamer: Temporal Lane Segment Topology Reasoning in Autonomous Driving

TL;DR

TopoStreamer tackles the problem of temporal lane topology reasoning for autonomous driving by introducing streaming attribute constraints, dynamic lane boundary positional encoding, and lane segment denoising. The method combines an BEV-based encoder with a dynamic, boundary-aware transformer decoder and a memory-based temporal propagation mechanism to maintain consistent lane segment attributes and topology across frames. Empirical results on OpenLane-V2 show state-of-the-art gains in lane segment perception and centerline topology, with ablations confirming the effectiveness of each proposed component and the importance of lossless streaming supervision. The work advances end-to-end temporal road-network understanding, with practical implications for robust planning and lane-change decision-making in dynamic driving scenarios.

Abstract

Lane segment topology reasoning constructs a comprehensive road network by capturing the topological relationships between lane segments and their semantic types. This enables end-to-end autonomous driving systems to perform road-dependent maneuvers such as turning and lane changing. However, the limitations in consistent positional embedding and temporal multiple attribute learning in existing methods hinder accurate roadnet reconstruction. To address these issues, we propose TopoStreamer, an end-to-end temporal perception model for lane segment topology reasoning. Specifically, TopoStreamer introduces three key improvements: streaming attribute constraints, dynamic lane boundary positional encoding, and lane segment denoising. The streaming attribute constraints enforce temporal consistency in both centerline and boundary coordinates, along with their classifications. Meanwhile, dynamic lane boundary positional encoding enhances the learning of up-to-date positional information within queries, while lane segment denoising helps capture diverse lane segment patterns, ultimately improving model performance. Additionally, we assess the accuracy of existing models using a lane boundary classification metric, which serves as a crucial measure for lane-changing scenarios in autonomous driving. On the OpenLane-V2 dataset, TopoStreamer demonstrates significant improvements over state-of-the-art methods, achieving substantial performance gains of +3.0% mAP in lane segment perception and +1.7% OLS in centerline perception tasks.

Figures (11)

• Figure 1: Comparsion between current streaming-based map learning methods yuan2024streammapnet and our TopoStreamer. TopoStreamer delivers more fine-grained road information through streaming perception of lane segments, which is vital for planning.
• Figure 1: The detailed architecture of decoder.
• Figure 2: The overall architecture of TopoStreamer. It consists of four main components: part (a) PV-to-BEV encoder for extracting BEV features from multi-view images, part (b) transformer-based decoder enhanced with a dynamic lane boundary positional encoding module to improve up-to-date positional information learning, part (c) DN query generator for lane segment denoising, and part (d) stream memory that enables temporal propagation.
• Figure 2: Qualitative results for lane segment denoising.
• Figure 3: Comparison of PE: (a) current streaming-based approaches yuan2024streammapnet, (b) recent single-frame detection methods li2023lanesegnet, and (c) our proposed method.