3D Object Recognition Of Scattered Parts Based On RealSense

Industry robot plays an important role in the field of automation,especially with widely use of information technology in the field of industrial automation.Intelligent technology of industry robot has raised widespread concerns in high-tech enterprises,institution and scientific research organization.And vision system is the key factor of robot grabbing systems.2D vision system has been a stable and common application in the industry.However,3D vision system is far from popular in practice as its low accuracy and speed.This project has designed a 3D visual recognition and positioning system,which is aimed at the robot grabbing systems.Besides,we built a software platform to carry out the experimental verification.A method to get the depth image of scattered parts based on a kind of depth camera named Real Sense SR300,is been proposed,and the point cloud data is obtained.By researching the imaging principle of Real Sense,the problem of noise or data missing in depth image can be solved by the filter algorithm based on Kalman algorithm or statistical method.The point data set obtained from a low-cost depth camera,would bring the problem in point cloud segment,especially in grabbing systems.In this paper,an easy and valid algorithm is proposed to solve this problem.With the method of octree,the filtered points set is divided into a number of voxel grid cubes.The cubes are next clustered into a number of sub-patch fragments by modified K-means algorithm,and all sub-patches will be finally joined independently to be a set of larger ones by two criteria,i.e.,extended convexity criterion and angle criterion,each of which corresponds to one working piece.With the global features and local features of point cloud,a method combining with histogram matching and local recognition algorithm of Bayesian is given to recognize the type of parts.Besides,by researching the principle and method of cloud registration,this paper estimate the parameters in matrix of rigid transformation by two steps,the result of first step is coarse and that of second step is exact.We get the matching pairs by matching the fast point feature histogram(FPFH)in two point cloud sets,and transformation matrix parameters is estimated coarsely by RANSAC method.Finally,the matrix parameters are optimized by Iterative Closest Point algorithm,and the transformation relationship between the target point cloud and the model point cloud is obtained,and the position of the target point cloud is deduced.An application software system has been developed based on Linux system,point cloud library(PCL)and Open CV and other open source platforms,and algorithm have been verified by many test cases,which have shown its potential applications in robot grabbing systems in prac tice.