A Lightweight Lane Detection Network Based on Adaptive Anchor Generation

In recent years, autonomous driving technology has become a global focus. With the aid of deep neural networks, lane detection technology has made significant progress; however, achieving accurate lane positioning remains a challenging task. Most previous work relies on predefined anchors, which somewhat limits the generalization of anchor-based methods in the field of lane detection, particularly in terms of flexibility. This paper proposes a new lightweight lane detection network based on adaptive anchor generation, named LAANet. In the framework, the Adaptive Anchor Generation Module (AAGM) predicts the starting point’s position and slope by analyzing the direction and contour of the lane lines in the image, generating anchors that are more closely aligned with the actual lane lines. The generated anchors are not only fewer in number compared to previous methods but also of higher quality than predefined anchors. To further aggregate contextual information, a Deformable Lane-aware Module (DLAM) that incorporates a channel attention mechanism is integrated to adaptively enhance useful features while suppressing irrelevant ones. This mechanism effectively eliminates redundant information, thereby further enhancing the feature representation of lane lines. Experimental results on the TuSimple dataset demonstrate that LAANet achieves competitive accuracy while maintaining a small model size of 41M.