Research On Industrial Object Recon-Struction,segmentation And Localiza-Tion Based On 3D Sensing Technology

Welding robots are relatively common robots in industry,which can help humans to be lib-erated from complex operating environments,improve labor productivity,and shorten production preparation cycles.In the welding production process,the accuracy of welding is affected by the accuracy of the workpiece produced in the previous process,the placement of the workpiece to be welded,and the welding seams deformation.Therefore,the positioning of the welding seams is an extremely important part.In the positioning of welding seams,coarse positioning and fine positioning are generally performed.Coarse positioning of the welding seams is the process of three-dimensional sensing by the robot.The coarse position of the object is determined by the perception of the environment,which prepares for the subsequent visual guidance of the welding seam tracking.The fine positioning of the welding seams is to determine the specific position of the welding seams using the image.3D reconstruction and point cloud segmentation are important technical components of 3D perception in coarse welding seam positioning.The accuracy of the former will be affected by cumulative errors,while the latter's commonly used learning method segmentation network is subject to uneven sampling of local feature extraction of point clouds.The effects of feature extraction networks are not deep enough.In the fine positioning of weld-ing seams,the existing calibration methods of welding robots guided by specific structured light plane vision systems have large errors and the need for high-precision calibration boards during calibration.In order to solve the above-mentioned problem of the position of welding seams,this paper studies the three-dimensional reconstruction of the welding seams,point cloud segmentation,and vision-guided fine positioning of the robot.The main contents are as follows:1.Aiming at the drift caused by the accumulated errors in dense 3D reconstruction,a loop detection bag of words model with correction of difference coefficients is introduced to perform trajectory correction to improve the performance of image similarity calculation.In addition,a method of entropy constrained camera pose to select key frames is proposed,which makes the frames with more severe camera movements easier to select as key frames,thereby improving the quality of loop detection.2.A deep learning network with point cloud instance segmentation is designed.The structure of resblock and denseblock is used to deepen the network.Multi-scale semantics are repeatedly aggregated in the hierarchical structure.The inverse density calculation of the network introduces the number of point clouds for kernel density bandwidth selection of the function.In addition,the high-dimensional features and low-dimensional features of the same local point cloud undergo similar dimensionality reduction learning after dimensionality reduction.Through the experi-ments on the segmentation of the public data set and the segmentation of the semantic scene,it is verified that the segmentation accuracy of the proposed network is improved.3.A hand-eye calibration method based on the structured light plane is proposed.The calcu-lation method of the hand-eye calibration matrix is obtained through mathematical derivation,and the bagging method is used to reduce the error.This method does not require customized high-precision calibration objects,and welding experiments have been performed in actual factories.It is verified that the hand-eye calibration using the bagging method can ensure the triaxial error within ±0.2mm,which improves the welding precision and quality of welding.