Recognition And Location Of Randomly Placed Workpieces Based On Laser Scanned Point Cloud

Due to the increase in labor cost and the diversification in product demand,robot is widely applied in industry.Consequentially it will be a necessary process for manufacturing transformation and upgrading to employ the multi-robot production line mainly composed of industrial robot.The introduction of three-dimension sensor(3D-sensor),enhances the spatial perception of the robot,which makes the production line system intelligent and skillful.Therefore,the research on the related theory in 3D-machine vision and the application in industry,has become the main focus of the domestic and international research.During the automatic assembly process,discernment and location of workpieces is a prerequisite for robot gripping and other related operation.In this paper,three-dimensional laser scanner is applied to the workpiece recognition and positioning system,and the detection and positioning method of target in scattered environment is studied.With a laser scanner,the scene spots clouds of random workpieces could be obtained,the estimation of workpiece pose can be carried on without dividing scene spots cloud.The main research is as follows:Firstly,the acquisition of the scene point cloud is studied based on the rotating laser scanner.By analyzing the principle of laser triangulation,this paper establishes the mathematical model of the rotating line laser scanning system.The composition of the system is given,the coordinate system is established,and the conversion relation of coordinate systems is studied.With the aid of a rotating laser scanner,the workpiece scene point cloud can be achieved to analyze the characteristics of point cloud data.Secondly,we study the pretreatment method of laser scanning point cloud.Considering the huge data of the laser scanning point cloud,the conditionally filtering algorithm and 3D voxel grid are adopt to simplify the point cloud data.The finding of k-nearest neighbors is also analyzed by virtue of the kd-tree to get the topology structure of point cloud.Besides,the cause of outlies is researched.Far and near outlies are filtered via the statistical filtering algorithm and SVLOF method,respectively.Further,the Principal Component Analysis is adopted to estimate the normal and curvature of point cloud.To solve the problem that the directions of normal are not consist,viewpoint manipulation and normal propagation method are used to adjust the normal of point cloud for scene and models.Thirdly,the extraction method of point cloud feature descriptor based on constraint is analyzed.The feature descriptor based points PPF and the local feature descriptor FPFH are studied respectively.According to the characters of scene point cloud,the features are filtered by setting curvature and distance threshold.Finally,research on target pose estimation based on Hoff weighted voting is conducted.According to the problem of target pose estimation based on Hoff weighted voting,partly feature-similarity weighting is adopted and the query of similar feature pairs are optimized by Hash table.Through the experimental platform,verifications of algorithm in this paper is proceeded by the simulated data and actual scan data,as well as the comparison with other algorithms.Under the same conditions,compared with the other algorithm,the proposed algorithm has no improvement in the detection time,but it improves the accuracy of recognition.