Airborne LiDAR Point Cloud Classification Fusing Dense Image Matching Point Clouds

At present,the classification method of airborne Li DAR point cloud by fusion optical image has been widely studied.However,most of such research focus on the spectral features of fusion image,while a few focus on the texture information and spatial information contained in the image.However,with the continuous progress and development of photogrammetry technologies,DIM point cloud data can be obtained by dense matching of aerial images.Some scholars have studied the point cloud obtained by image matching,but there are few studies on the fusion of airborne Li DAR point cloud and DIM point cloud.In order to better use the spatial information of the point cloud generated by image matching,a multi-Tr Adaboost is proposed in this paper to fuse the features of airborne Li DAR point cloud and DIM point cloud,which can effectively improve the classification accuracy of airborne Li DAR point cloud.At the same time,the proposed classification method uses only the geometric features of samples,so it is not necessary to register DIM point cloud and airborne Li DAR point cloud in samples.Through the improved multi-Tr Adaboost model,the airborne Li DAR point cloud and DIM point cloud in different regions can be integrated to reduce the requirements of DIM point cloud measurement area and make full use of the existing aerial images.The specific experimental contents are as follows:(1)Data preprocessing: according to the difference between airborne Li DAR point cloud and image dense matching point cloud,Cloud Compare was used to thin out the airborne Li DAR point cloud with high density,and the DIM point cloud with high impurity was de-noised.Then the software is used to classify the two kinds of point clouds and the classification results were used as the evaluation criteria of the experiment.(2)Collection and extraction of the best features: in this paper,the multi-scale geometric features of point clouds are selected for classification without considering the color information.Moreover,the random forest algorithm is used to rank the importance of the features,and then the best selection is made.In addition to single point,point cloud segmentation was also carried out based on voxel and super-voxel,and the best selection of voxel and super-voxel was extracted.(3)Construction of multi-Tr Adaboost model: the Tr Adaboost algorithm of traditional binary classification is improved,making it suitable for multi-category point cloud classification.Three experiments were carried out on the basis of point,voxel and super-voxel to prove the effectiveness of the proposed method.The multi-Tr Adaboost model proposed in this paper can automatically assign weight to DIM point cloud and realize the purpose of fusion of DIM point cloud to airborne Li DAR point cloud classification.The experimental results show that the multi-Tr Adaboost model proposed in this paper can effectively utilize the spatial information of DIM point cloud and improve the classification accuracy of airborne Li DAR point cloud,which is particularly advantageous in the classification of vegetation points.The fusion of DIM point clouds in different regions can also play a very good auxiliary effect,mainly reflected in the improvement of the classification accuracy of building points,which proves that the method proposed in this paper has good universality.