The Research On Pose Estimation Method Of Mechanical Parts Based On 3D Vision

6D pose estimation technology is an important part of robot intelligence,and the accurate and fast 6D pose estimation algorithm can improve the robot’s perception of the external environment.At present,the 6D pose estimation of target objects is mainly for desktop-level objects,which is quite different from the real production environment in the industry.Weak or untextured target objects,complex lighting conditions,and severe mutual occlusion all pose higher challenges for 6D pose estimation.In this paper,the 6D pose estimation method of weak texture or untextured mechanical parts is studied from the actual situation and demand of industrial production.The specific contents are as follows:(1)In order to make the data more consistent with the actual situation of industrial production,this paper selects six kinds of mechanical parts as research objects,and uses the Blenderproc method to make a simulation data set of mechanical parts,which provides data support for subsequent algorithm research.(2)Due to the poor quality of the point cloud restored according to the depth map,in order to improve the quality of the point cloud data,this paper uses the radius filtering method to process the point cloud,remove redundant background points and noise,and indirectly improve the accuracy of 6D pose estimation.In this paper,the curvature downsampling method is also used to select the key points of the point cloud,which avoids the problems of voids or missing features caused by the random downsampling method,and can make more effective use of the geometric features of the target object.(3)According to the existing 6D pose estimation algorithm,reliable RGB features cannot be extracted when facing weak texture or no texture objects,resulting in the problem of accuracy degradation.In this paper,it is proposed to add the point cloud normal information of the target object to the 6D pose estimation method,and enhance the adaptability of the 6D pose estimation network in the face of different objects through the adaptive weighted multi-feature fusion network based on attention mechanism.It enables good pose estimation results whether it is facing rich textured objects or weak textured or untextured objects.Experimental results on Linemod dataset and self-made simulation dataset show that the proposed method has higher accuracy than other 6D pose estimation methods,which confirms the effectiveness of the proposed method.(4)Aiming at a large number of standard parts and common parts in the mechanical industry,this paper studies the category-level 6D pose estimation method.The feature extraction ability of the target instance is enhanced by bidirectional fusion feature network,and the category-level 6D pose estimation of industrial standard parts is realized by combining the category shape prior network.Through experiments on NOCS dataset and self-made simulation dataset,the proposed method achieves good results in the comparison of different indicators,which confirms the effectiveness of the proposed method.