Segmentation And Classification Of 3D Point Clouds Of Outdoor Scenes

At present,3D point clouds are widely used in intelligent unmanned vehicles and object detection.Segmentation and classification of 3D point clouds are two important research directions in the computer vision field,especially for outdoor scenes.Being able to accurately detect and identify objects in outdoor scenes is of great significance for hot tasks such as intelligent vehicles driving and land autonomous navigation.Point clouds collected by 3D laser scanners in outdoor environments are a set of sparse,massive and disordered points.All the points in the set constitute an outdoor scene.The segmentation and classification of 3D point clouds of outdoor scenes is a challenging task due to the complexity of the outdoor scenes.In this paper,the neighborhood characteristics and topological structures of 3D point cloud data of outdoor scenes are studied deeply,and the reasonable segmentation and classification of 3D point cloud data of outdoor scenes are realized.In the aspect of ground extraction,the main innovation of this paper is to propose a ground extraction method based on the regression of the two-dimensional gaussian process.Firstly,an initial ground result is extracted quickly with the improved incremental algorithm,and then the accuracy of the ground extraction is further improved with the two-dimensional gaussian process regression method.The ground processed by the two-dimensional gaussian process regression method is taken as the final ground.After the ground points are extracted,the distance between the non-ground points is clearer and the change in density is more obvious.This greatly reduces the difficulty of subsequent object segmentation.In terms of segmentation and classification,this paper adopts the DBSCAN algorithm for unsupervised clustering of non-ground points.This algorithm can divide regions with sufficient density into clusters and realize arbitrary clustering in 3D spatial data with noise points.Finally,each cluster of point clouds is classified by means of the PointNet neural network.This neural network requires that every sample in the training set must contain the same number of 3D points.Since objects have different sizes,and the numbers of 3D points in different objects vary greatly.Therefore,when we make the neural network training data set,the improved lattice method is used in this paper to align the numbers of 3D points constituting various objects,so that each training sample contains the same number of 3D points.