| Road marking,as one of the most basic road elements,is a research objective bringing large of attentions.The accurate extraction of road markings is of great significance in the development of advanced driving assistance system and high precision map.With the advantages of insensitivity to light conditions,vehicle laser point clouds can provide abundant information of 3D space and reflection intensity,to lay as an indispensable data source for automatic extraction of road markings.However,since the point clouds have uneven distribution on reflection intensity and density or low contrast between road line and its surrounding road surface,the existing thresholding method is difficult to extract road line accurately.It is necessary to implement study on road marking extraction method to improve its precision based on vehicle LiDAR point clouds.The deep learning method is highlighted by depth features adaptively from shallow features by using neural networks,which has been received wide attentions to target detection,image classification,etc.This study introduces deep learning techniques to explore new approach for road marking information extraction based on vehicle LiDAR point clouds.From the perspective of semantic segmentation deep learning of this study gives the following contents:(1)The "DL road marking extraction dataset based on MLPC" was established.Since there is no open-source dataset for road marking extraction by using point cloud reflection intensity information,the vehicle-mounted laser point cloud data was selected and preprocessed,followed by converting the data into intensity feature images.The road marking information was labeled according to the road marking-related specifications,combined with the markings involved in the original vehicle-mounted point cloud data.The methods of data chunking and segmentation,as well as data enhancement and normalization were studied.(2)To address the issue related to unbalanced samples of road marking area and road background area,resulting in failure on common loss functions that cannot accurately describe model errors,a new loss function,DC Loss was constructed by integration of the loss functions Cross Entropy Loss and Dice Loss.Such function improved the training effect of the deep learning model and improved the accuracy of road marking extraction.(3)An attention-aware SA-Unet-based road marking extraction method was proposed.In order to enhance the feature extraction capability of the U-Net model with the improved loss function,the road marking extraction method based on the attention-aware SA-Unet was constructed by introducing the mixed-wash attention-aware module SA-Net.Such improvement promotes interaction of channel spatial information,further to enhance the model ability to screen road markings and road background.The results show that,(1)the "DL road marking extraction dataset based on MLPC" is successfully constructed;(2)the DC Loss function constructed is able to improve the training effect of the model compared with other common loss functions,to achieve a higher accuracy of road marking extraction.(3)The introduction of a mixed-wash attention-aware module SA-Net can help the network to perceive accurately the channel space feature relationship between road markings and road background,thus to improve the efficiency and accuracy of markings extraction.(4)Compared with other thresholding-based road marking extraction algorithms such as adaptive thresholding,the proposed method is highlighted by robustness and generalization,which may overcome the extraction issues related to unever distribution of reflection intensity and point intensity.The road marking extraction method based on attention-aware SA-Unet has better fittness results,which provides insight into the development and application of high-precision maps for autonomous driving.This study also lays out foundation for improvement on the precision of road marking extraction from in-vehicle LiDAR point clouds. |
PDF Full Text Request