3D Object Recognition Method For Industrial Assembly Parts

Object positioning is a prerequisite for robot operations such as sorting,loading and unloading,spraying,welding,and assembly.Although the traditional structured positioning method using fixtures as a carrier has the characteristics of good reliability and high precision,but the design process is complicated and non-universal,and can’t meet the needs of flexible and intelligent manufacturing.With the development of machine vision,robots have been given"eyes",enabling robots to use visual sensors to obtain information,thereby improving the level of robot intelligence.However,visual information is related to various factors such as object material,texture,environment,etc.At present,more and more industrial fields use robots for sorting,assembling and other operations.According to the type of images obtained by the sensor,it can be divided into object recognition based on two-dimensional images and threedimensional images.Compared with two-dimensional images,three-dimensional images contain more information and are less susceptible to the effects of illumination and object stripes.The six-degree-of-freedom pose of the object can be obtained directly from the threedimensional image.It is more suitable for mechanical parts recognition with single color and unobvious texture features.By comparing the existing four types of 3D object recognition methods,it is found that the recognition method based on local features does not need to segment the target point cloud,and can be applied to the object recognition with the target point cloud shaded.However,the existing methods have some shortcomings in terms of stability and recognition efficiency.Therefore,this thesis starts from the above two aspects.An efficient and stable 3D object recognition algorithm based on local features is studied.Firstly,in order to improve the efficiency of 3D object recognition algorithm,this thesis proposes E-SAC-IA algorithm based on the commonly used point cloud initial registration algorithm SAC-IA,which is improved from two aspects of sampling target and sampling criteria.One is to change the sampling target from point to point pair to reduce repetitive calculation;Secondly,a similar,triangle filtering mechanism is proposed to filter out the sampling points that cannot produce the correct solution to improve the sampling efficiency.Experiments show that compared with SAC-IA,the efficiency of object recognition based on E-SAC-IA algorithm is improved tens of times.In addition,sampling point constraint is added on the E-SAC-IA algorithm,and a multi-object recognition algorithm of 3D point cloud is proposed and verified.Secondly,in order to improve the stability of 3D object recognition algorithm,from the aspect of keypoints,this thesis proposes a key point extraction algorithm based on the curvature information of point cloud.Firstly,calculating the curvature of the preselected points in the point cloud,and the initial keypoints are selected according to the curvature value.Finally,the local non-maximum suppression is carried out,which using the curvature value to obtain the final keypoints of the point cloud.Experiments show that the performance of the keypoints extraction algorithm proposed in this thesis is better than the existing several commonly used keypoints extraction algorithms,and the object recognition success rate based on this algorithm is higher,all above 90%.Since the model point cloud in this thesis is generated by using 3D modeling software to draw CAD and turn it into point cloud,in order to obtain the point cloud of the model more easily,a simple and quick way to obtain the point cloud of the 3D CAD model by interactive reconstruction of a handheld object is studied,the method combined with RGB image and the depth map,and manipulating images by using the hand information,thereby obtaining Multiview point clouds of object,Finally,the 3D CAD model point cloud is generated by registration and merge multi-frame point clouds.Finally,in order to verify the accuracy and feasibility of the 3D object recognition algorithm proposed in this article.Obtaining the workpiece posture by this study algorithm and uising hand-eye calibration parameters converting workpiece posture to hand claw grasping posture,thereby grabbing the workpiece by ABB 1200 robots.The experimental results show that recogniting the workpiece position based on this study algorithm is accurate,and the experimental scheme is feasible.