SMFRNet: Complex Scene Lane Detection With Start Point-Guided Multi-Dimensional Feature Refinement

Lane lines play a crucial role in the traffic system. However, due to the diversity of lane categories, road conditions and weather environments, as well as the different aspect ratios of lane lines, lane detection algorithms face many challenges. This paper proposes a multi-dimensional feature refinement method for complex scene lane detection based on start point guidance. Due to the fact that lane lines often traverse the entire image, capturing sufficient context is crucial, and lane lines also have specific local patterns that require detailed low-level features for accurate localization. We propose global feature refinement (GFR) and lane aware gather (LAG) to refine the features from the following two dimensions: global enhancement and local refinement. To generate high-quality anchors, we predict the start point coordinate of lane instances through start point coordinate prediction (SPCP). To better fit 2D lane detection, we adopt the more general penalty LaneIoU (PLIoU) as the loss function to evaluate the predicted results. Experimental results demonstrate that the proposed method performs well in in lane detection tasks in complex scenes and has strong competitiveness among existing methods.