Enhancing Lane Segment Perception and Topology Reasoning with Crowdsourcing Trajectory Priors

TL;DR

This work tackles robust online lane-level mapping by introducing trajectory priors derived from crowdsourced data to augment lane segment perception and topology reasoning. The authors encode trajectories into two priors—Rasterized Heatmaps and Vectorized Instance Tokens—and fuse them with BEV features through a BEVFormer-based encoder and a DETR-based map decoder, guided by an alignment module to address spatial and semantic misalignment. Extensive experiments on OpenLane-V2 show that trajectory priors significantly improve $AP_{ls}$ and $TOP_{lsls}$, outperforming state-of-the-art methods and demonstrating strong cross-scenario generalization. The approach offers a practical pathway to richer, up-to-date lane topology information for autonomous driving systems and HD-map construction.

Abstract

In autonomous driving, recent advances in lane segment perception provide autonomous vehicles with a comprehensive understanding of driving scenarios. Moreover, incorporating prior information input into such perception model represents an effective approach to ensure the robustness and accuracy. However, utilizing diverse sources of prior information still faces three key challenges: the acquisition of high-quality prior information, alignment between prior and online perception, efficient integration. To address these issues, we investigate prior augmentation from a novel perspective of trajectory priors. In this paper, we initially extract crowdsourcing trajectory data from Argoverse2 motion forecasting dataset and encode trajectory data into rasterized heatmap and vectorized instance tokens, then we incorporate such prior information into the online mapping model through different ways. Besides, with the purpose of mitigating the misalignment between prior and online perception, we design a confidence-based fusion module that takes alignment into account during the fusion process. We conduct extensive experiments on OpenLane-V2 dataset. The results indicate that our method's performance significantly outperforms the current state-of-the-art methods. Code is released is at https://github.com/wowlza/TrajTopo

Figures (4)

• Figure 1: Comparison of Different Priors: a) Ground Truth. b) SDMap, where blue lines represent roads, green lines represent sidewalks, and red lines represent crosswalks. c) Trajectory. The trajectory data is more closely match the geometric structure of lanes, especially in the intersection area.
• Figure 2: The overview model architecture. The model receives surrounding images, a local aligned SDMap and trajectory data as prior inputs and the model follows the typical encoder-decoder paradigm. We add the sdmap upon previous work. Then we innovatively represent trajectory data in two forms and compare different fusion strategies, ultimately identifying the optimal fusion strategy.
• Figure 3: Illustration of alignment Module: We initially concatenate the prior feature with BEV feature and feed them into several convolutional layers to predict the coordinate offsets at each position. Subsequently, we employ a warp operation to achieve spatially aligned feature. These aligned feature then proceed through the confidence fusion module to further integrate the prior information. Throughout the training process, BEV segmentation supervision is implemented to correct misalignments in both space ans semantics.
• Figure 4: Qualitative results:It can be observed that under certain complex road conditions, the coarse sdmap provides some prior information, but its simple structure does not guarantee that the model has a more comprehensive understanding of the road scenario. In contrast, trajectory priors, due to their closer alignment with real-world scenarios, achieve better supplementary effects