Intelligent Labeling Of Three-dimensional Point Cloud Data

The application of three-dimensional point cloud in actual production and living has become more and more extensive with the development of corresponding technologies,such as pattern recognition,reverse engineering,and autonomous navigation.In the process of processing point clouds,the marking and picking operations of point clouds are very important.By picking up the point cloud data,a series of operations such as addition,deletion,modification,and check of the point cloud can be completed.At the same time,the most important thing is to add a category tag to the selected point cloud by selecting and marking the point cloud.The point cloud tags in this paper are mainly divided into three types of tagging methods: artificial point cloud tags,segmented point cloud tags,and smart point cloud tags.The traditional marking method,namely artificial marking,has high accuracy and good effect,and can modify the marked scene in real time.In this paper,the 3D world coordinate system is transformed into the matrix coordinate system of the model view through the model view matrix,and the coordinates are projected by establishing the cube in the coordinate system established by the model view matrix.The projection matrix transformation method is used to obtain the coordinates of the projected point,and then the coordinates of the vertex are mapped to the [0,1] interval.Finally,the screen coordinates after the conversion of the threedimensional point coordinate value are obtained through the viewport and position calculation.After obtaining the screen coordinates,use OpenGL rendering and MFC message mapping to perform mouse and keyboard control to select the area to be marked.The advantage of the artificial marking method in practical application is that the marking process is convenient,rapid,and accurate.The disadvantage is that in the process of marking,it can be found that there is a situation in which easy marking errors occur during the process of selecting a point cloud,and in the face of a large number of scenes of point clouds,manual marking takes time and effort.So this article has improved the point cloud marking.After roughly dividing the scene,mark it again.Doing so can reduce the complexity of the scene structure,and divide the scene into independent point cloud collections,which is convenient for operators to pick up and mark part of the point cloud,and realize conversion from a single point cloud object to a class point cloud object.In the segmentation process of the point cloud,firstly,the covariance matrix is established to obtain the feature values and feature vectors of the point cloud.By judging and comparing the feature values,the point clouds in the scene are divided into three categories,and the three types of vectors are utilized.Through the establishment of feature balls and Meanshift clustering algorithm,the point cloud data is normal vector clustered.Then using DBSCAN algorithm to point cloud data location.Then,the relationship between the average normal vector and the average elevation and the threshold of each type of plane point cloud after clustering is determined to obtain a preliminary ground.If the vertical distance from the preliminary ground to the ground point near the preliminary ground and the point of the non-preliminary ground is less than the preset threshold,it is merged to the preliminary ground to obtain the final ground.After the final ground is obtained,the ground-removing part is further segmented using the DBSCAN algorithm to complete the segmentation process of the outdoor three-dimensional point cloud scene.Marking the segmented scene can convert a single point into a type of point cloud tag.Although marking the segmented scene greatly simplifies the operation complexity of the marker,the efficiency of the marker is improved.However,the accuracy of the mark has a large dependency on the accuracy of the segmentation result of the point cloud,and the marking of the segmented scene still requires the intervention of the operator.This paper proposes to build a random forest model classifier by using training set samples to intelligently mark unknown scenes.This kind of labeling method gets rid of the operator's intervention and makes the mark more intelligent.In the process of smart tagging,the basis for the node splitting of the CART tree and the selection of conditions for stopping the splitting of the CART tree are mainly defined.After smart tagging of point cloud data is completed,the result of point cloud smart tagging caused by node splitting and termination splitting conditions selection or data complexity may be corrected by manual marking.The operator can manually mark the unknown scene according to the smart tagged scene of the point cloud.In this way,the function of the unified cooperation between the artificial marker and the smart marker point cloud is realized.In practical applications,the scene needs to be marked according to the complexity of the scene.