Research On Identification And Pose Estimation Method Of Untextured Metal Workpiece

In recent years,the robot industry has developed rapidly,and the manufacturing industry has become increasingly intelligent.In the trend of gradually releasing human resources,the mechanical arm has gradually replaced human beings to do complicated and difficult or more delicate work.As the working environment of the manipulator becomes more and more complex,the scenes in the practical application environment are chaotic,the target objects and other similar objects are stacked,and the position and posture of the target are complex and changeable,which cause great troubles for the grasping and picking tasks of the manipulator.Therefore,the 6D pose estimation of the research object is of great research significance for the accurate completion of these tasks.Therefore,pose estimation of research object has great research significance.In this thesis,the pose estimation of weakly textured workpiece is taken as the purpose,and the messy scene of workpiece is designed for data acquisition.The endto-end pose estimation method based on voting is studied and grasping verification is carried out.The contents are as follows:（1）In terms of data acquisition,label Fusion is introduced to collect real data of weakly textured artifacts because the common data set does not meet the problem of weak textured artifacts in industrial scenes,which can retain the semantic correlation between different frames at the maximum cost.Due to the small size of self-made data set,different data enhancement methods were explored to improve the robustness of the model,and successfully input into the network to achieve good results.（2）In the aspect of 6D pose estimation,the attention mechanism is integrated into the pixel-level voting network for pose estimation.Since the empty convolutional network cannot fully utilize all pixels in the image in the feature extraction process,which is fatal in the pose estimation method based solely on RGB,the network is designed as a standard deep convolutional network,and the attention mechanism is introduced for feature extraction.In view of the weak texture characteristics of the workpiece,and other workpiece in the same scene belong to the same material and the same surface,the improved voting network is used for key point detection,compared with before improvement,using Add-s evaluation criteriaim proved 1.8% on artifacts dataset.When exploring whether the use of deep networks for smallscale data will be over-fitting,the networks of Res Net-18 and Res Net-50 with different residual units are used as the backbone networks for pose estimation tasks.（3）Based on the constructed visual manipulator system,the 6D pose estimation method of the workpiece was verified experimentally,and the proposed network was used to conduct pose evaluation experiment on the common data set Line Mod,which achieved good visualization results.Finally,the object in the self-made data set was taken as an example to realize the grasping task of the workpiece.