Semantic Segmentation Of Low-Quality Point Clouds Based On Deep Learning

Point cloud is a collection of points in a three-dimensional space.It is often noisy or even missing due to the influence of the scanning instrument and the shooting environment,forming the low-quality point cloud.Three-dimensional point cloud segmentation technology is an important part of scene understanding.It is often used in autonomous driving and SLAM technology.It can segment different categories in realtime and accurately in scenes containing pedestrians,building road surfaces and other objects.At now,although three-dimensional point cloud segmentation method has been developed to a certain extent,because low-quality point cloud models usually have noise and missing,the robustness and accuracy of current point cloud segmentation models for low-quality point clouds can be further improved.This thesis systematically analyzes the existing problems and main ideas of the current point cloud segmentation methods,and on this basis,combined with deep learning methods,proposes an effective solution to the problem of low-quality point cloud segmentation.At the same time,in order to make up for the lack of 3D human body point cloud data set,a low-quality 3D clothing human body data set was established.Therefore,the main research contents of this article include:1.This thesis studies and analyzes the current point cloud segmentation algorithm and its existing problems,and summarizes the main ideas and existing problems of various algorithms.Discusses the advantages and disadvantages of traditional mathematical methods and point cloud segmentation methods based on deep learning.This thesis focuses on point cloud segmentation methods based on deep learning,analyzes the limitations of these methods for low-quality problems,and makes an evaluation on the data set proposed in this thesis.2.This thesis proposes a point cloud data segmentation algorithm based on overlapping region retrieval and alignment,and discusses and analyzes some of the problems to be solved in the algorithm design in detail: the retrieval based on overlapping regions,the detection of overlapping regions,the optimization of overlapping regions,the alignment of overlapping regions,and the design of loss function.At the same time,it compared with the current commonly used point cloud segmentation algorithms,and discussed the practicability of low-quality and large-scene point clouds.3.In order to solve the new human-computer interaction system,the human body model has gradually become the key issue of human-computer interaction,and at the same time,to make up for the blank of the three-dimensional clothing data set,and to further promote the development of point cloud segmentation technology.This article collects a three-dimensional human body clothing data set and compares it with some existing human body data sets.Laid a data foundation for the algorithm research proposed in this thesis.Finally,a prospect is made for further research work,which provides some important ideas for follow-up research.