Semantics Based On Local Features Of 3D Point Clouds Research On Segmentation And Its Application

With the advent of the information age trend,the research in the field of artificial intelligence has become an irresistible trend.How to identify the object in a complex scene is a very important research content in deep learning.Point cloud semantic segmentation is an important means to analyze and understand 3D scene,which has broad application prospects and research value in many fields.In recent years,with the rapid development of deep learning and 3D point cloud data acquisition technology,the direct semantic segmentation of point cloud data using deep learning method has attracted more attention.The method based on deep learning can extract the deep semantic information of point cloud and mine the important features between data.However,at present,many deep neural networks do not make better use of the local information between point clouds,and are unable to deal with large-scale point cloud data and partial missing of point cloud data,resulting in low segmentation accuracy.Aiming at the above problems,this paper focuses on the extraction of local features of 3D point cloud,aiming to propose a semantic segmentation network of multi-level point cloud based on effective local feature extraction module,and optimize and improve the problem of missing point cloud.The main research work is as follows:(1)Construction of local feature extraction module.For large-scale disordered point cloud data,this paper establish Kd-tree data structure organization,using KNN algorithm for each point ou recent K neighborhood point in space,by local area code for each point add redundant information,will be the output of neighborhood characteristics of polymerization to obtain each point characteristics after the aggregation,improve the feelings of each point.(2)The construction of semantic segmentation network.In this paper,we follow the encoder decoder architecture,stack multiple local feature extraction modules to significantly increase the receptive field of each point.By analyzing the characteristics of FPS,IDIS and RS sampling methods,we use RS sampling method to down sample the large-scale point cloud data,extract the point cloud feature information hierarchically,and splice it through the jump cascade way to integrate the multi-resolution feature information Information,make full use of the features of shallow and deep network,enrich the features learned by the whole network,and improve the accuracy of semantic category prediction.(3)Point cloud completion and sampling mode adjustment.Considering the influence of partial missing of point cloud data on semantic segmentation of the whole point cloud,a method of point cloud data completion is proposed.First,the boundary of an object is extracted,all internal hole boundaries are found and points are evenly scattered to fill the internal hole.At the same time,various clustering methods are analyzed,and an improved sampling method based on k-means clustering is proposed.By using k-means clustering different characteristic information of point cloud for many times,the point cloud data sampled by the network is more uniform.(4)In order to verify the overall performance of the point cloud segmentation network proposed in this paper and the influence of the improved method on the semantic segmentation results,a series of training and tests were carried out on the S3DIS data set.Success results show that the proposed network is a good performance in relatively ideal segmentation effect,and at the same time,it verifies that the improved method proposed in this paper can improve the semantic segmentation results.