Object Recognition And Pose Estimation For Bin Picking

As the "Holy Grail" problem of manipulator operation,Bin Picking has a wide range of application scenarios.Stable and fast object recognition and pose estimation algorithms are one of the keys to Bin Pickin problem.In this paper,aiming at the Bin Picking scene,the sample synthesis method based on object rendering is used.Based on the CAD model of the known object,a large amount of training data is synthesized to construct a virtual dataset,and the Mask R-CNN network is used to realize the multi-workpiece recognition and segmentation of the stacked scene.After obtaining the single-target workpiece data,an improved algorithm based on point pair features is proposed.For the problem of poor real-time performance of traditional algorithms,a densefusion network method based on deep learning is proposed.The specific research work of this article is as follows:(1)Aiming at the self-occlusion problem of the object itself,the traditional point pair fea-ture algorithm does not consider the problem of invisible point pairs.This paper proposes an improved PPF algorithm.The algorithm includes two processes,offline training and online matching.In the offline training phase,spherical Fibonacci sampling is used to obtain the adopted viewpoint,and the visibility judgment is performed by the HPR algorithm under mul-tiple viewpoints.The visible points that meet the conditions are combined in pairs,and the point pair features is calculated to obtain a model that describes the global information of the object.In the online matching stage,the farthest point sampling algorithm is used to obtain the reference points in the scene.The reference points and the local point cloud are composed of point pairs to calculate features,and Hough voting is used for rough matching.Finally,the ICP algorithm is used to fine-tune the pose results.This paper conducts experiments on the UWA dataset and actual collected point clouds,and conducts a robotic grasping experiment on a Bin Picking experiment platform.The results verify the effectiveness of the proposed method.(2)Aiming at the problem that the PPF algorithm requires a large number of feature match-ing and Hough voting,and the real-time performance is insufficient,this paper proposes a dense-fusion network based on RGB-D data.The densefusion network combines two data sources of image and point cloud,and uses PSPNet and PointNet to extract deep features from RGB images and point clouds,and perform pixel-level dense fusion to obtain multi-scale features.In the ro-tation prediction module,the 20-hedron group is used to sample the rotation anchor points,and each anchor point is predicted,which reduces the influence of the local optimization problem.In the translation prediction module,according to the displacement vector predicted by each pixel,the RANSAC method is used to predict the translation transformation,so as to obtain the precise pose of the target object.The sample synthesis method based on object rendering provides a large amount of labeled data for the network,which greatly reduces the cost of data annotation.In this paper,experiments are carried out on the LINEMOD data set,synthetic pipe data set and real scenes,and the results verify the effectiveness of the method.