Research On Perceptual Algorithms In 3D Large-scale Point Cloud Scenes

With the advent of new concepts such as autopilot,high-precision maps and smart cities,many scenarios require perception of 3D environment perception and interaction based on point cloud.The research of perceptual algorithms in large-scale point cloud scenes has a wide range of application prospects.Point cloud is an important 3D geometric data structure,which can accurately and directly reflect the real world.Therefore,3D point cloud data is used as a carrier to be studied three scene-perception challenge tasks:classification,point cloud segmentation,and 3D object detection in largescale point cloud scenes.In the task of classification in 3D large-scale point cloud scenes,considering the influence of uneven sampling,sensor accuracy and other factors,two point cloud filtering algorithms for large-scale scenarios are proposed to improve data quality.Because point cloud has a high degree of sparseness and disorder in space,a new local geometric relationship operator,which is named Graph Embedded Network,for extracting point cloud space is proposed,it links the characteristics of the point cloud structure with the graph.The similarity between points is used as the basis for 3D feature selection to strengthen the local features of the point cloud.In order to comprehensively describe the characteristics of point clouds,according to the point cloud spatial distribution,14 different point cloud observation perspectives are set to project point cloud to supplement the point cloud 2D feature information,and strengthen the local feature expression ability of point cloud classification networks.Based on the 2D and 3D features of the point cloud,a point cloud classification network with composite visual features is designed,and the effectiveness and robustness of the algorithm are verified on the ModelNet40 dataset.In the task of segmentation in 3D large-scale point cloud scenes,the limitations of traditional 3D point cloud segmentation algorithm are analyzed first.Therefore,in order to retain the geometric information of point cloud as much as possible,a encoding—decoding structure,which is named Pyramid Attention Network is designed to enhance the strong semantic features of each point while increasing the receptive field.Then Combined with the Graph Embedded Network,a 3D point cloud semantic segmentation network framework is designed.The framework is suitable for two point cloud segmentation tasks: Part segmentation and semantic segmentation.In the task of 3D object detection in large-scale point cloud scene,the classification network and the segmentation network are combined to form a new point cloud processing structure textit pyramnet,to enhance the expression ability of point cloud features.The 3D object detection framework encodes features in the column direction of point cloud space,and applies textit pyramnet to extract multi-scale features.Secondly,a new 3D region proposal network(3D Region Proposal Network,3D RPN)is designed to objects with different aspect ratio by downsampling in different directions.Finally,the validity and feasibility of the network structure are proved on the Kitti verification set.In summary,this thesis combines traditional algorithms and deep learning for processing three basic scene-aware tasks in large-scale point cloud scene,and verifies the validity and feasibility of the above studies on authoritative data sets.