| With the continuous promotion of urbanization in China,rail transit has become the main means of rapid urbanized transportation due to its unique advantages of high efficiency,large capacity,and high space utilization.To improve the safety operation coefficient of rail transit train operation,the real-time detection of obstacles in front of the track area of the road network becomes especially important.The existing obstacle detection scheme based on visible light sensors lacks depth information and is susceptible to ambient light conditions.In view of the stable 3D position information acquisition capability of the LiDAR,this paper focuses on the 3D point cloud semantic parsing method for the complex environment of rail transportation to improve the accuracy of obstacle detection.However,the difficulties of 3D point cloud semantic parsing are(1)the huge amount of manual annotation;(2)the lack of sufficient texture information to guide the automatic annotation.Therefore,in this paper,by fusing RGB image data from a 2D camera,the annotation results obtained from image semantic segmentation are migrated to point cloud data to achieve automatic annotation of 3D point clouds;then,based on the point cloud annotation data of the rail transportation environment,the point cloud semantic segmentation model is optimized to find the balance between low latency perception capability and high accuracy detection performance.The main contributions of this paper are as follows.(1)A monocular ranging algorithm for rail traffic scenes.This paper uses the special condition that the track lines are parallel and the spacing is known to achieve monocular ranging based on 2D images for any point on the line,so that the depth information of other objects of unknown size in the scene can be accurately estimated.The method does not require detection and recognition of objects of known size to complete the estimation of scene depth,and the method can be extended to any scenes with parallel relationships.The algorithm performs well in actual working scenes with a range accuracy of 98.10%.(2)Automatic external calibration algorithm based on camera and LiDAR.By solving the depth of image pixel points,the depth information of two-dimensional pixel coordinates is supplemented,and the joint calibration problem of multiple sensors is reduced to the calculation of the translational and rotational positional matrices,so as to realize the calculation of external parameters by using the three-dimensional line correspondence relationship under two coordinate systems.The method can be calibrated online and in a single frame,saving a lot of manpower.The calibration error value of its single-frame data is 3.24 pixels,which can be reduced to 2.25 pixels after multiple-frame iterations.(3)Semantic annotation method for 3D point clouds incorporating image semantic parsing.This paper migrates the two-dimensional image semantic annotation to the threedimensional point cloud semantic annotation,and uses the more mature image semantic segmentation technique to obtain the results of environment understanding and analysis,and then migrates them to the point cloud,which is a faster and more convenient point cloud annotation method with high accuracy.(4)An improved semantic segmentation network framework for point clouds.Based on the existing network architecture RandLA-Net,which is excellent for large-scale scene segmentation,the residual network is directly connected,the feature dimension is increased and recompressed by sharing MLP,and the important information is selected and saved,to increase the ability of the method to adapt to low arithmetic constraints and achieve further improvement of accuracy.The model can reach 78.5%of mloU in the standard dataset Semantic3D(reduced-8),and 82.4%in the actual established orbital field attraction cloud dataset. |
