Multi-level Classification Of Large-scale Airborne Laser Scanning Point Clouds

As an active measurement method,ALS technology has the ability of all-day and all-climate observation,as well as the ability of vegetation penetration.It has advantages in the rapid acquisition of large-scale 3D geographic data,and has been widely used in forestry survey,basic surveying,power inspection and other fields.Compared with the rapid development of hardware systems,the intelligent processing technology of point clouds is relatively backward.The classification of massive ALS point clouds consumes lots of effort,time and money in real-world applications.To address this problem,scholars at home and abroad have conducted lots of in-depth research on the automatic classification of ALS point clouds,and achieved certain results.However,the existing methods are mostly limited to specific scenes and lowlevel hand-crafted features,which can not meet the accuracy and stability required for practical applications.In recent years,deep learning has broken through the difficulty of traditional machine learning methods relying too much on hand-crafted features,and has achieved great success in various image recognition tasks.In this case,to better realize the automatic classification of 3D point clouds,this paper combines deep learning technology to carry out in-depth research on scene recognition,filtering and fine-grained classification of ALS point clouds.The main research contents are as follows:(1)ALS point cloud Scene recognition can provide scene prior information for other 3D classification tasks such as point cloud filtering.However,there is currently a lack of systematic research in this area.To address this problem,a point cloud scene recognition method based on deep fusion of multi-view and multimodal feature maps is firstly proposed.To facilitate the direct utilization of the existing mature 2D CNNs and their models pre-trained on large-scale image datasets,a multi-view and multimodal feature map mapping method of point cloud is proposed.The method maps a point cloud into a series of 2D multimodal feature maps along multiple perspectives.Compared to single view,multiple views can more fully encode the 3D information of the point cloud.In addition,to better fuse multi-view feature maps,a two-level fusion network is designed.The network embeds both feature-level and decision-level fusion strategies.The comparison experiments on the dataset show that the proposed scene recognition method is better than the other three adapted classical recognition methods,and can accurately identify nine types of 3D terrain scenes.(2)ALS point cloud filtering is the key to efficient acquisition of high-precision DEM.Due to the low efficiency of existing point cloud filtering method based on deep learning,a point cloud filtering method based on variable resolution voxel grid and sparse convolution is proposed.To make the filtering model more flexible to deal with complex and varied terrain scenes,a variable resolution voxel grid generation method of point cloud is proposed.The method first performs ring compression on the space outside the center of the original point cloud,and then regularizes the compressed point cloud into a variable resolution voxel grid.The variable resolution voxel grid can cover more spatial context information while ensuring a high resolution in the central region.In addition,to process large sized voxel grids and adopt large network structures,a 3D semantic segmentation network based on submanifold sparse convolution is constructed.The comparison experiments on the dataset show that the proposed point cloud filtering method is much more efficient than the existing point cloud filtering method based on deep learning.At the same time,it is superior to existing commercial software in terms of terrain feature preservation and automation.(3)ALS point cloud fine-grained classification is the basis for applications such as fine-grained 3D reconstruction,power corridor inspection,and 3D high-precision map production.Existing point cloud classification methods based on deep learning often use only a single structure network or single modal input,and have insufficient ability to distinguish similar and complex geometric objects.To address this problem,a point cloud fine-grained classification method based on 2D-3D integrated CNN is proposed.To improve the ability of the model to distinguish geometrically similar objects,a 3D CNN with multimodal fusion mechanism is designed.To avoid the pregeneration step of the feature map,a 2D multi-view CNN capable of internal 3D to 2D transformation is designed.In addition,to eliminate the local noise existing in the classification results of the integrated CNN,a post-processing method based on fully connected CRF is proposed.The comparison experiments on the ISPRS public dataset show that the average F1 value and the average class accuracy achieved by the proposed point cloud fine-grained classification method surpass the previous best results,respectively.The multi-level classification methods of large-scale ALS point clouds proposed in this paper will contribute to improve the intelligent level of Li DAR point cloud data processing and promote the application of ALS technology in more fields.