SDLane: Efficient Lane Detection via Sparse Anchors and Dynamic Fusion
TL;DR
SDane, a highly efficient, accurate, and concise lane detection model, is proposed that achieves high operational speed and F1-score while using significantly fewer line anchors than comparable methods.
Abstract
The complexity of real-world scenes poses significant challenges to lane detection. Existing methods often sacrifice inference speed and model simplicity by relying on excessive anchors or complex associations. To tackle these issues, we propose SDLane, a highly efficient, accurate, and concise lane detection model. First, we establish sparse labels, named SparseLabel, by selecting representative directional points. Then, we design a sparse lane anchor generation module. Based on the sparse direction map and confidence map predicted by this module, the number of anchors required for detection can be reduced from hundreds or thousands to a quantity comparable to the number of lane instances. These high-quality line anchors enable SDLane to operate without relying on computationally intensive feature extraction or fusion methods such as FPN or attention mechanisms, thereby boosting operational speed. Furthermore, the designed dynamic kernel fusion module enhances the fusion of sampled features for lane anchors while achieving a lightweight and efficient design through low-rank decomposition. In addition, we construct a dedicated lane detection dataset RainyLane for testing in heavy rain environments. Experiments on four datasets, encompassing challenging environments such as curved roads, blizzards, rainstorms, occlusions, and nighttime, demonstrate that SDLane achieves high operational speed and F1-score while using significantly fewer line anchors than comparable methods. For instance, on the CULane test set, SDLane achieves an F1-score of 79.76% and a speed of 486 FPS using only five anchors. The code is available at: https://github.com/happy-q123/SDLane .