Automatic Lane Marking Extraction From Mobile LiDAR Point Cloud by Integrating BEV Projection and Sparse Polyline Annotation

Abstract

With the rapid development of intelligent driving and highdefinition maps, the demand for accurate and efficient automatic lane marking extraction is increasing. However, existing methods are often heavily dependent on scene-specific features, limited robustness is exhibited in complex road environments, and sufficient semantic information for lane representation is often lacking. To address this issue, the structured representation of point cloud data and the topological modeling of lane markings is investigated, and a lane extraction method integrating bird’s-eye view (BEV) representation with sparse polyline annotation is proposed. First, a BEV construction method that integrates trajectory constraints with multiattribute collaborative encoding is developed. Second, a sparse polyline annotation system that integrates endpoint positions, directions, instances labels, and semantic attributes is designed. Finally, a hybrid feature learning framework combining convolutional neural networks (CNN) and vision transformers (ViT) is constructed. Compared with conventional BEV-based lane detection frameworks, it enables joint learning of local geometric details structures and long-range topological dependencies, thereby improving the robustness and structural consistency of lane marking extraction. To evaluate the performance of the proposed method under diverse road scenarios, the model is trained using multiscenario data, including highways and urban roads, while the test set consists of independent of the training set urban roads and rural lanes characteristics. Experimental results demonstrate that this method outperforms existing approaches in lane coordinate prediction, endpoint localization, and semantic segmentation. Specifically, the F1-score for lane coordinate prediction exceeds 91%, while the F1-scores for endpoint localization and semantic segmentation both exceed 82%.