The paper proposes a novel lane detection framework named ElasticLaneNet that models lanes as zero-contours on a flexibly shaped Elastic Lane Map (ELM). The key highlights are:
ELM representation: Lanes are implicitly represented as zero-level contour lines of ELM, which can flexibly capture diverse lane geometries including large curvature, intersections, Y-shapes, and dense lanes.
Elastic Interaction Energy (EIE) loss: The EIE loss function guides the training of ELM, enabling the model to consider the global context and low-level features. The long-range attractive interaction in EIE helps the predicted lanes converge to the ground truth, even under weak lane features.
Network architecture: ElasticLaneNet is built upon an Encoder-Transformer-FPN backbone, with Elastic Lane Map Module (ELMM) as the lane detection head. Auxiliary modules like Transformer Bottleneck and Feature Fusion are incorporated to improve performance.
Experiments: ElasticLaneNet achieves state-of-the-art results on the structurally diverse SDLane dataset, outperforming existing methods in complex scenarios like dense lanes, large turns, and Y-shaped lanes. It also shows competitive performance on the TuSimple and CULane datasets.
The paper demonstrates that the geometry-flexible ELM representation and the EIE loss-guided training enable ElasticLaneNet to effectively handle challenging lane detection cases while maintaining efficient inference speed.
To Another Language
from source content
arxiv.org
Дополнительные вопросы