Research On Pose Estimation Of Stacked Workpieces Based On 3D Vision Technology

In recent years,the achievements created by artificial intelligence technology have achieved great success,especially in the industrial field,where industrial manufacturing is changing from manual work to intelligence.The industrial assembly line still requires workers to place the workpieces neatly or mechanically separate the stacked workpieces,which is in urgent need of intelligent transformation.Workpieces always interfere with each other in stacked scenes,so the features of workpieces are very incomplete and the poses of different workpieces are quite different,which gives a huge challenge to the recognition and positioning capabilities of the vision system.In addition,scenes with weak texture,stacked workpieces,and reflection appear frequently in industrial environments,and 2D vision technology is difficult to solve the problem well.For the actual need of industrial assembly,a set of pose estimation scheme for stacked workpieces is designed based on 3D vision technology and experimental verification is carried out.In order to overcome the lack of data sets in stacked scenes and the high cost of manual labeling,a simulation data set generating and automatic labeling process is designed.The scenes with stacked workpieces can be simulated through the simulation environment.Point cloud of the scene is recovered from the depth image captured by virtual camera and the data set is automatically annotated.The volume of the simulation data set increases through data augmentation,and the simulation data can be more similar to the real data.The whole process is simple and fast,with low cost,and can provide lots of point cloud data.In order to simplify the stacked scenes,to provide point cloud information of a single workpiece,a point cloud instance segmentation network is built.The feature extraction network is used to extract the features of the point cloud,and three-dimensional bounding box is used to constrain the position of the point cloud.A small number of predicted bounding boxes are set and the boundary coordinates of the bounding box are directly regressed.By scoring the mask of each point,the correspondence between the point and the instance number is realized.A multi-criteria loss function is designed to supervise the training of network,and only the simulation data set is used to complete the training and tuning of the point cloud instance segmentation network.Based on the segmented instance point cloud,the accurate pose estimation of a single workpiece point cloud is realized.The key point selection,feature extraction,and feature matching are performed on the point cloud,and the matching relationship between the point cloud of model and the point cloud of scene is obtained.The pose estimation problem is transformed into a truncated least squares optimization problem based on matched points.The TEASER algorithm is used to solve the optimization problem,and the ICP registration algorithm is used to optimize the pose estimation result.The experimental platform is built,and on this basis,the experiments of point cloud instance segmentation,pose estimation using point cloud,and the robotic arm grasping are carried out to verify the proposed pose estimation scheme for stacked workpieces.The results of experiments show that the constructed point cloud instance segmentation network can generate valid segmentation results for scenes with different numbers of workpieces,different point cloud resolutions,and unknown objects.In real scenes,the pose estimation algorithm using point cloud has an average translation error of 0.4 mm,an average rotation error of 0.75°,and a grasping success rate of more than 90% for stacked workpieces,so it can meet the requirements of automatic grasping in scenes with stacked workpieces.