Semantic Classification Of Outdoor Scenic Spot Pointcloud Data For Tree Feature Extraction

3D laser scanning is the key technology to obtain outdoor 3D spatial information quickly,and large-scale point cloud classification plays an important role in autopilot,intelligent transportation,augmented reality,forest remote sensing and other fields.In particular,with the national attention to environmental protection and ecology,laser scanning technology in forest resources survey and plant ecology and other fields are more and more widely used,through the laser point cloud can accurately estimate the biophysical characteristics of trees,accurate and efficient access to the location of single wood,leaf and wood points and parameter information.A prerequisite in these applications is how to properly classify and target segment large-scale point clouds.However,there are difficulties such as large scale,complex and diverse scenes,uneven density,overlapping masks between targets,and noise interference,which bring great challenges to the classification of large-scale outdoor scene point cloud data.And the tree canopy overlap,staggered distribution,leaves and stem point clouds are seriously affected by the quality of laser scanning,which brings challenges to its intelligent processing.Therefore,this paper focuses on the scientific problem of how to accurately and efficiently classify and divide in large-scale point cloud scene,from the theoretical,key technology and so on.The specific contents of the study and the results achieved are as follows:(1)In the traditional point cloud classification,the feature engineering-based method relies on the design and selection of artificial features.The feature calculation method is time-consuming and has insufficient expression ability,and needs to set a large number of parameters.Its classification results are greatly affected by the experience of researchers,therefore,this paper based on deep neural network method to achieve point cloud classification.Due to the large amount of point cloud data,the point cloud data is diluted based on the random subsampling method to improve the algorithm efficiency,and the local context information is fused to enhance the ability of feature expression.Secondly,the attention mechanism is used to obtain more representative information,and the problem of overfitting is avoided by residual connection when the original spatial resolution is recovered by upsampling.In this way,the classification of outdoor point cloud data is realized,which lays a foundation for subsequent tree segmentation and branch and leaf classification.(2)In this paper,density-based point convolution and Pointnet feature forward propagation layer are firstly combined to embed tree point cloud into high-dimensional instance space and semantic space.Secondly,the neighborhood information of points in the instance space is searched,and the semantic information is fused to enhance the feature expression.The instance information is fused with the representative semantic information extracted based on the graph attention mechanism.Finally,the highprecision segmentation of the tree point cloud is realized based on metric learning method.(3)Due to the influence of laser scanning quality,it is difficult to distinguish the branch points from the leaf points.In this paper,the results of tree point cloud segmentation are used to classify branches and leaves of each tree.Firstly,according to the linear distribution of trunk points and scattering distribution of leaf points,the original point cloud information with distinctive geometric features is calculated.Secondly,based on anisotropic convolution,the spatial context features of the point cloud are transformed to solve the disorder problem of the point cloud.Finally,the spatial resolution of original point cloud was recovered by the nearest neighbor up sampling to achieve the separation of branches and leaves.On the basis of branch and leaf classification,the trunk point cloud was extracted,and point cloud slicing was done at the DBH measurement place.Then,the DBH parameters of trees were extracted by fitting the slicing data based on RANSAC method.This paper uses the public Semantic3 D,TUM,NPM3 D data sets and roadside trees,forest trees data sets to experimentally verify the proposed algorithm.The attention neural network that integrates local context information has made some progress in semantic labeling of large-scale outdoor scenes.The average accuracy and average intersection ratio reached 92% and 76.8%,respectively,and the classification accuracy of trees reached 96%,which is accurate The tree point cloud is extracted from the ground;the average segmentation accuracy of the single tree segmentation method combining tree point cloud instances and semantics reaches about 86%,which can effectively segment trees with close canopy intervals;the average classification accuracy of the branch and leaf classification method based on deep learning reaches97.67%,the accuracy is equivalent to or better than the most advanced branch and leaf classification method in the previous research,which can effectively distinguish the branch and leaf points,and better solve the problem of easy misdivision of branch and leaf points.