3D Point Cloud Segmentation In Different Scenes

In recent years,with the consistent development of 3D scanning technology,the processing algorithm for 3D point cloud has gradually become a research hotspot.Compared with the traditional 2D image,the 3D point cloud contains more spatial structure and feature information,and it is easier for the machine to understand the scene.In order to understand the scene information correctly,we segmentant the point cloud into different objects in the scene to lay the foundation for subsequent processing.The segmentation of 3D point cloud is a very important part of the automation tasks such as object classification,recognition,positioning and navigation.The result of segmentation will affect the next processing task directly.This paper mainly studies and compares the mainstream point cloud segmentation algorithm.To solve the problems in indoor and outdoor scenes,we propose some improvements in these two scenarios.Then we validate the results by experiments.The main work of this paper is as follows:(1)Firstly,we introduce some common methods and research status of point cloud segmentation.We also introduce the application scenarios of point cloud segmentation and point cloud preprocessing methods,such as point cloud filtering and point cloud registration.This is the basis for point cloud processing.The super-voxel clustering method of point cloud is introduced in detail,which lays a foundation for an improved segmentation algorithm based on local convexity.(2)In the scene of indoor segmentation,the main problems are as follows.:(a)If the objects in the scene are similar in shape and close to each other,it is difficult to divided;(b)If there are many noise in the scene,the traditional method of regional growth will be greatly affected by noise which leads to mis-segmentation.Therefore,this paper introduces an improved point cloud segmentation method based on local convexity and dimension features.Some improvements have been made to the selection of seed points and the growth rules.We use RANSAC algorithm to remove the desktop and other planes in the scene,and we conpare the original algorithm and the algorithm by experiments.The results show that the improved algorithm can reduce mis-segmentation to some extent.(3)In the scene of outdoor segmentation,the processing speed of point cloud is relatively high.Due to the large amount of calculation,it is difficult to meet the requirements of real-time.Aiming at this problem,and combined with the particularity of the outdoor scene,we propose an outdoor scene point cloud segmentation algorithm.This method firstly removes the ground point cloud data by an improved RANSAC method,and then divides the remaining data by clustering.Since the RANSAC algorithm is random sampling,non-ground points are often sampled which causing unnecessary calculations.We use the characteristics of ground point cloud data to select the vicinity of the radar sensor and the height is lower than the fixed value data as the sample,to reduce the number of iterations to improve the processing efficiency.Through experimental verification,the processing time of each frame is about 60ms,which can basically meet the requirements of real-time.