Road markings are important traffic sign information,and onboard LiDAR point clouds provide high precision 3D co-ordinates and reflectance intensity information for their extraction.Due to factors such as scanning distance and target material,the different object may exhibit similar intensity values,causing interference in the extraction of road markings.Wear and aging during road use can also damage the original structure of the markings,resulting in discontinuities after extraction.In addition,the diversity of road markings and their varying occurrence frequencies in practice can lead to low classification accuracy for categories with fewer samples in the segmentation network extraction results.To address these issues,this paper proposes a two-stage segmentation and classification extraction method that accurately extracts various types of markings and has topological robustness.Firstly,a multi-layer perceptron is used to adaptively learn the relationship between intensity and its influencing factors,and to perform intensity correction on the road point clouds.Secondly,the semantic segmentation network link spatial topology net(LST-Net)is proposed to segment all road markings,which captures line structure information using row-column convolution and attention mechanisms,and is trained with topological punishment to determine the positions of markings.Finally,YOLOv5 is used to detect the markings,and a separate classification network is trained to address the issue of sample imbalance in segmentation.Experiments are conducted on three sets of point clouds from different driving scenarios,and the results show that our approach achieves a marking extraction accuracy of 94.1％and a recall rate of 95.6％,demonstrating strong practicality and effectiveness.
vehicle-mounted LiDAR point cloudsroad markings extractiontopological structuresemantic segmentationobject detection
万方数据
维普
