The Research On Deep Learning Method Of Object Pose Estimation Based On Point Cloud Model

Three-dimensional point cloud is an important data structure that can provide richer geometric information than images.With the continuous improvement of sensing technology and the continuous breakthrough of deep learning in the 3D field,the deep learning method based on point cloud has become a brand-new solution in many industrial automation applications.Object pose estimation is the key to many scene perception applications,and point cloud data that is not sensitive to lighting and texture information gradually plays an important role in it.In this paper,focusing on the specific application of robot grabbing randomly,aiming at stacked industrial parts in unstructured environment,a deep learning method of object pose estimation based on point cloud model is proposed.The main content includes the following four parts:(1)Aiming at the problem of excessive time consumption in labeling real dataset,a simulation dataset generation method is proposed,which uses physics engines to simulate the real scene with stacked industrial parts,and renders the scene point cloud.The points in each sample contain corresponding semantic labels and instance labels for segmentation training,and each instance contains corresponding pose label for training of point cloud pose estimation.(2)A point cloud semantic instance segmentation network is proposed,which can perform semantic segmentation and instance segmentation on scene point clouds to obtain single instance point clouds for subsequent pose estimation.This paper improves the fusion method of semantic features and instance features,which significantly reduces the memory consumption of the network.At the same time,it uses two clustering algorithms to obtain purer instance point clouds.Experiments show that the proposed segmentation network can accurately identify scene instances,and the models in the simulation dataset can be well migrated to the real dataset.(3)A projection-based point cloud pose estimation network is proposed,which can predict the accurate 6D pose of a single input point cloud.In this paper,the point cloud is projected onto a two-dimensional plane to generate multiple flat feature maps that are easy to handle,and merge them to significantly improve the accuracy of pose estimation.At the same time,aiming at the ambiguity of symmetrical parts,a weighted loss function is proposed.Experimental results show that the pose estimation method is better than the traditional method based on feature matching in terms of stability and speed.(4)For the proposed two-stage pose estimation network,a robot binpicking system platform is built to verify the effectiveness of the algorithm in an unstructured environment.This paper directly uses Python script to write integrated programs,establishes communication with the robot through Socket,and controls the robot to perform grab operations.The entire algorithm flow is visualized through Open3 D.Experiments show that the point cloud pose estimation method proposed in this paper is still highly robust in complex scenes with occlusion and stacking,and can successfully complete the task of industrial parts sorting.