Research On 6D Pose Estimation Method For Scattered Texture-less Workpieces

6D pose estimation of texture-less workpieces in highly cluttered scenes is a key and challenging problem in the field of visual perception.In order to solve the problem of insufficient precision of 6D pose estimation of texture-less workpieces,this paper studies the algorithm of 6D pose estimation and the algorithm of pose refinement based on RGBD,and proposes the method of 6D pose estimation based on position dependent adaptive fusion and the method of pose refinement based on surface normal.The main contents of this paper are as follows:(1)Aiming at the problem of lacking workpiece dataset,the method of making dataset is studied.A data acquisition system is built by using a Yaskawa robot and a depth camera.Yaskawa robot is used to control the movement of depth camera and collect the data of target workpieces.Then the scene is reconstructed,and the pose of the target workpiece in the scene is annotated by using the key point matching and ICP algorithm.(2)In order to solve the problem that the precision of 6D pose estimation for texture-less workpieces is insufficient due to the lack of feature information,this paper proposes a 6D pose estimation method based on position dependent adaptive fusion(PDAFusion method).In this algorithm,the position dependent feature extraction module is introduced to extract the position dependent relationship between each pixel to enrich the feature information of each pixel,so as to improve the accuracy of each pixel in predicting the pose and the confidence.At the same time,an adaptive feature fusion module is introduced to obtain the weight of the feature information of different data sources in each pixel,so as to solve the problem that the feature information of different data sources has different contributions to the 6D pose estimation task.The experimental results show that the accuracy of the PDAFusion method proposed in this paper is improved by 1.3% on YCB＿Video dataset,0.6%on Linemod dataset and 1.1% on the self-made workpieces dataset compared with other methods.(3)Aiming at the problem of low time efficiency or insufficient feature information in pose refinement method,this paper proposes a pose refinement method based on surface normal(SNPRefine method).The algorithm further improves the 6D pose estimation accuracy of weak texture workpiece by introducing surface normals.Surface normals are extracted from the depth data of the target object by using the surface normals acquisition module,and the normal feature information of the weak texture workpiece is obtained through a convolutional neural network with an encoder-decoder archtitecture,so as to enrich the global features and improve the accuracy of the pose refinement network.The experimental results show that the accuracy of the SNPRefine method proposed in this paper is improved by more than 1% either on the public datasets YCB＿Video and Linemod or on the self-made workpieces dataset,and the time cost is lower than ICP algorithm.Aiming at the problem of insufficient feature information on the color images of scattered and texture-less workpieces,this paper studies the 6D pose estimation and pose refinement algorithms of the workpiece,and proposes the 6D pose estimation method of PDAFusion and the SNPRefine method,which improve the 6D pose estimation accuracy of texture-less workpieces,and have a certain contribution to the realization of the intelligentization of industrial robots.However,in the future,further research on the 6D pose estimation method for thin and elongated objects with weak texture is needed to further improve the generalization performance of the 6D pose estimation method.