Research On Point Clouds Registration Algorithm Of Scattered Parts And Application Of Robot Grasping

Nowadays,the manufacturing industry of China has entered the ' Industry 4.0 ' era with intelligent manufacturing as the core,and the demand for automatic robot grasping is growing.The difficulty of automatic grasping technology is to identify the position and attitude of the parts(pose of the parts),and the most convenient and intuitive method in the recognition of pose is the registration algorithm of 3D point clouds.Therefore,it is of great theoretical and practical significance to study the point clouds registration algorithm in the recognition of pose for scattered parts in the industrial field.this paper aims at the integrity of point cloud of scattered parts,the point cloud registration algorithms of regular and losing are studied respectively,and the practical application research is carried out combined with robot grasping.The main research contents of this paper are as follows :1.Studying the object segmentation method of scattered parts point cloud.Firstly,the voxel grid algorithm used to simplify the acquisition of point cloud,the statistical filtering algorithm used to remove the outlier noise points and the random sampling consistency algorithm used to remove the scene noise points is to complete the preprocessing.Then,the segmentation algorithm based on region growing is used to realize the complete segmentation of a single part,and the stability of the segmentation result is further improved by using the mothed of the point cloud comparison based on offline template and the mothed of the Z-axis numerical comparison of the centroid.2.Proposing a point cloud registration algorithm based on adaptive principal component analysis aim at regular point cloud.The principal vector of the point cloud is obtained by the principal component analysis algorithm,which combined the spatial characteristics of the point cloud and the Cartesian coordinate system rule to adaptively adjust the direction of the vector.Then,it is combined with the fine alignment to complete the pose recognition.Experiments show that under the same accuracy,the registration speed is nearly 4 times faster than the registration algorithm based on the direction histogram signature feature,and has great real-time advantages.3.Proposing a point cloud registration algorithm based on extended clustering Gaussian image aim at losing point cloud.Firstly,Gaussian mapping is performed on the point cloud,and the obtained extended Gaussian image is clustered and segmented.Then,the volume-norm operator is used for matching,and the optimal corresponding sub-point cloud is found and the inverse Gaussian mapping is performed.Finally,the distance-curvature descriptor is used to query corresponding point pairs realized the coarse registration,and the pose recognition is completed by combining the fine registration.Experimental results show that under the same accuracy,compared with the registration algorithm based on direction histogram signature feature,the registration speed of scattered rod parts point cloud with local missing is also increased by nearly32.5 %.4.Combining the regular and missing two types of point clouds,a combined optimal registration method is proposed.Firstly,the adaptive principal component analysis algorithm use for coarse registration and the volume ratio of its bounding box is used as its registration error to judge the registration error.If it fails,the extended Gaussian image clustering registration algorithm continued used for the registration result.Finally,the registration process of the overall system is completed by combined with fine registration.So,it makes the overall system with the best real-time performance and stability.5.Using the combined optimal cloud registration method,the robot grasping of scattered parts is studied.The single point of multi-pose principle is used to calibrate the tool grasping point of the robot,and then the hand-eye calibration is completed by using multiple sets of pose data of the same object in the same ' hand-eye ' relationship.In order to select the optimal grasping path point,a multi-point template calibration method is proposed,and the grasping pose is solved by combining the point cloud registration algorithm.Using the three-dimensional scanner,industrial robot,industrial computer and other equipment,a complete set of three-dimensional visual robot grasping hardware system are built.Based on Visual Studio 2017 development platform and PCL1.8.0point cloud library,a three-dimensional point cloud poses recognition software is developed.The grasping experiment shows that under the premise of ensuring the accuracy of pose recognition,the pose recognition time of the system is not more than 2s,which can meet the real-time and stability requirements of general industrial robots for grasping scattered parts.