RBCANet: Recursive and Bidimensional Context Aggregation Network for Lane Detection

Authors:

Abstract
Lane detection plays an important role in current autonomous driving systems. Even in scenarios lacking visual guidance, such as adverse weather conditions or low-light environments, the ability to detect lanes remains essential. In this paper, we present the Recursive and Bidimensional Context Aggregation Network (RBCANet) aiming at enhancing detection accuracy when lane lines are not visible. Specifically, we introduce the RCA module, which leverages convolutional neural networks and attention mechanisms to collect local and global information respectively, effectively aggregating contextual information from the feature map and establishing long-range dependency. Additionally, we present the BCA module, further enhancing the feature representation of lane lines through two branches: channel attention and spatial attention. Quantitative and qualitative experiments conducted on the TuSimple and CULane datasets demonstrate that our proposed RBCANet achieves excellent detection performance.