🤖 AI Summary
This work addresses the high cost and infrequent update challenges of high-definition (HD) maps, as well as the limited reliability of purely sensor-based online mapping in long-range and occluded scenarios. To this end, the authors propose D2HDMap, a system that leverages a lightweight, imperceptible driveline prior to guide an end-to-end neural network for online vectorized HD map prediction. By incorporating a noise-aware training strategy, the method effectively fuses prior knowledge with real-time perception. Remarkably, using only this low-cost and easily maintainable driveline prior, D2HDMap significantly enhances model generalization even in the absence of such priors during inference. Experimental results demonstrate that D2HDMap achieves 44.8 mAP on geographically disjoint test sets of nuScenes and Argoverse 2, outperforming state-of-the-art methods while exhibiting greater robustness to localization errors.
📝 Abstract
Accurate, up-to-date representations of road structures are critical for the safe operation of autonomous vehicles. Existing systems rely either on costly, maintenance-heavy high-definition (HD) maps which compromise safety when outdated, or purely sensor-based online mapping which struggles with long-range reliability and occlusion. Systems incorporating map prior information into online mapping seek to overcome drawbacks of both approaches by combining them in some way. We propose 'Driveline To HD Map' (D2HDMap), an online mapping system that injects a lightweight, non-visible driveline prior to guide the estimation of visible road structures such as lane dividers, road boundaries and crosswalks. This prior incurs less effort to create and update compared to full HD map priors used in other approaches. We also show that training with such a prior can improve generalization at inference time when no prior is available. Ablation studies conducted on the nuScenes and Argoverse 2 dataset demonstrate that models trained using a driveline prior largely retain performance even when priors are not available. On a geographically disjoint split, D2HDMap achieves 44.8 mAP, surpassing recent state-of-the-art. Additionally, noise-aware training substantially increases robustness to realistic localization error.