Research And Application Of 3D Point Cloud Recognition Algorithm For Stacked Parts

The machine learning based goal recognition method complements the classification and recognition of the goal in the scene by extracting and learning the test functions.With the rapid development of machine vision theory,the research has gradually shifted from independent object recognition to stacked target recognition.The efficiency and accuracy of the stack target recognition algorithm based on point cloud plays a significant role in the realization of machine vision system,and is the basis of guiding the robot to complete the complex motion of parts identification,sorting and so on in the industrial field.The thesis focuses on the goal recognition of three-dimensional point cloud.The main research work are as follows:（1）The k neighborhood search is realized by using kd-tree algorithm,and the normal vector and curvature information of point cloud are estimated by principal component analysis.In order to suppress the influence of point cloud noise on the recognition accuracy,a method of denoising the point cloud filter by combining through filtering and voxel grid down-sampling is proposed.The experimental results show that the Bunny model point cloud can be reduced from 35947 points to 18782 points,and the outliers and noise points in the original point cloud can be effectively filtered out.（2）To solve the problem of difficulty of defining the number of clusters and poor precision of target objects in pose clustering,a three-dimensional point cloud recognition algorithm based on point-pair features and hierarchical complete-linkage clustering is proposed.Point pair feature descriptors are used to extract features,and local reference frame is introduced to transform the field and model point clouds.Hough transform voting and hierarchical complete-linkage clustering are used to obtain the initial pose of the target point cloud.Finally,the precise registration of the pose of the workpiece is completed by iterative closest algorithm.The comparative experiment shows that the clustering contour coefficient of the algorithm is closer to 1,the degree of aggregation is smaller,and the clustering effect is good.（3）For the purpose of verifying the validity and accuracy of the proposed algorithm,target recognition experiments are carried out using the common data sets of Stanford University and the University of Western Australia and the point clouds of the actual collected stacked parts.The results show that the target object recognition efficiency is higher when the relative spacing coefficient θ=0.1 is applied.The values F₁ of different methods in different scenes are all within the range of[0,1],indicating that the proposed algorithm has a good recognition effect.