Multi-target Segmentation And Edge Extraction Based On 3D Point Cloud

In recent years,due to the unique spatial structure advantage of 3D point cloud,it has been widely used in the fields of surveying and mapping,medical treatment,artificial intelligence,leisure and entertainment,etc.However,because the collected data is easily affected by factors such as camera resolution,mechanical vibration,and collection environment,the data is subject to redundancy and noise interference.Therefore,in order to make the 3D point cloud better used in various fields,high-quality denoising,effective segmentation and efficient edge extraction of the original point cloud data obtained by the scanning device are very critical parts.In this context,based on the data collected by the TOF camera,the multi-object segmentation and edge extraction of the 3D point cloud are studied.The main research content of this paper includes the following aspects.This paper studies the random sampling consistency plane extraction algorithm in complex scenes.The algorithm only depends on 3 points to obtain the model coefficients,which has great instability.This paper proposes an adaptive random sampling consensus algorithm based on PCA for this problem.In the improved algorithm,first of all,N point clouds in the local area are selected at random,and the covariance matrix about these N points is constructed by principal component analysis.According to the eigenvalues derived from decomposition of the covariance matrix,the regional distribution characteristics of N point clouds are judged.Repeat the above process until the samples distributed in the same plane are selected.Secondly,according to the feature vector corresponding to the minimum feature value,the initial plane model is fitted.Then,an adaptive threshold is used to obtain the internal points.Finally,by iterating the above steps,the plane model with the largest number of internal points is selected.Compared with the original plane extraction algorithm,the proposed algorithm can extract the plane model more accurately,the average error rate is only 2.83%,and the accuracy rate is 97.17%.After removing background noise,effective information needs to be obtained by target segmentation.In this paper,the Contractive thinking and cosine similarity are introduced into the Euclidean segmentation algorithm,and an improved multi-objective segmentation algorithm is proposed.According to the characteristics of the experimental data,the point cloud data is contracted along the normal vector to enlarge the Euclidean distance between the targets,and the cosine similarity weighted Euclidean segmentation algorithm is used to segment the contracted point cloud data.Compared with the Euclidean segmentation algorithm,the improved algorithm has better performance on the adjacent and occluded boundaries,and each target can be completely segmented,the target segmentation error rate is only 3.25%.Edge information plays a vital role in the measurement of some object sizes.Therefore,in order to improve the efficiency of the edge extraction algorithm based on the angle threshold,this paper proposes a grid-based secondary edge extraction algorithm.A large number of non-edge points are removed by grid point processing and point cloud density statistics to obtain initial edge candidate points,and then the final edge data is obtained by the maximum angle threshold algorithm.In the experimental environment of this paper,point cloud data test is carried out on the algorithm.The improved algorithm can quickly extract edge feature points while ensuring complete target contour points.The efficiency of the improved algorithm is 21.26% higher than the grid and angle threshold fusion algorithm.