Refinecurvelane: lane detection with B-spline curve in a layer-by-layer refinement manner

Wei Tian, Yi Han, Yuyao Huang, Xianwang Yu

Abstract: Lane detection with front-view RGB images has been a long-standing challenge. Among the various methods, curve-based approaches are known for their fast speed, conciseness, and ability to handle occlusions. However, these methods often suffer from a relative low accuracy, attributing to the inflexibility of adopted curve model, the inefficient lane feature extraction, and a rigid curve regression supervision. In this paper, we propose a novel curve-based lane detection method that addresses these limitations. The lane lines are modeled with B-splines, which provide greater flexibility. Explicit spatial attention maps are used to guide the network in extracting relevant lane features from the image. Additionally, a layer-by-layer refinement process is employed to improve the lane predictions. Importantly, the ground truth of spatial attention maps also serve as pixel-level supervision for the lane instances. We evaluate the proposed method on four widely used lane detection datasets and demonstrate the state-of-the-art performance achieved among curve-based approaches on CULane and LLAMAS dataset.