Research On 3D Point Cloud Scene Segmentation Method

In recent years,with the rapid development of 3D scanning technology,the acquisition of3 D point cloud data has become more and more convenient.Compared with two-dimensional images,three-dimensional point cloud data can express more information,and the point cloud data is less affected by external environmental factors,such as illumination,rotation,etc.,which makes more researches turn to three-dimensional point cloud data.Point cloud segmentation is to segment different objects in the scene to realize the semantics of the scene.In the complex and changeable external environment,point cloud segmentation can help the machine to better understand the scene and perceive the intelligent environment,which is essential for tasks such as object positioning,recognition,classification and feature extraction.After studying and analyzing the current situation of point cloud segmentation,aiming at the problems existing in the current deep learning and traditional methods in point cloud segmentation,the existing methods are improved.Firstly,this paper briefly introduces the research background and significance of point cloud segmentation technology,application prospects and research status at home and abroad.Then the basic concept of point cloud and data preprocessing operation are described.Finally,some common techniques of point cloud segmentation based on traditional methods and deep learning are introduced respectively,which will lay the foundation for the subsequent work of point cloud segmentation.Aiming at the current point cloud deep learning network's insufficient description of local geometric figures and the interaction between points,and the inability to make full use of remote context information,the paper designs a neural network that can be directly used for 3D point cloud data.The network can learn to enrich the local geometric relationship and the relevance of context features.By designing the local feature extraction module,the local neighborhood is constructed by searching the neighborhood points in multiple directions.The local neighborhood features are summarized by learning the neighborhood point relationship,and then the center point feature is obtained by using the improved pooling method.After the feature extraction module,the spatial attention module is used to learn the correlation between any two points in the feature map to further improve the expression ability of features.Aiming at the problems of the traditional point cloud segmentation,such as the difficulty of filtering ground points and the slow speed of obstacle segmentation,this paper designs a point cloud segmentation algorithm based on the traditional method.First,through an effective mapping method,the 3D point cloud is projected into the 2D space,and the data is gradually partitioned,and a simple line fitting method is designed in each partition to complete the extraction of ground points;Then,the angle threshold algorithm is used to segment the non ground points quickly on the point cloud depth image.The points marked as not belonging to the ground are grouped to complete the object segmentation.Finally,collision detection is carried out for the segmentation results.Finally,this paper uses the Scan Net and Shape Net data sets to compare the point cloud segmentation based on deep learning,design ablation experiments,noise experiments and random data loss experiments to prove the feasibility and robustness of the network.In the point cloud segmentation algorithm in outdoor scenes,comparative experiments on the design of ground filtering algorithm under different road conditions have verified that the algorithm in this paper has a good ground filtering effect.Through the comparison of different segmentation methods,the effectiveness of the non-ground segmentation algorithm is demonstrated.