Research On 6D Object Pose Estimation Method Based On Topological Features And Pose Decoupling

With the increasing development of intelligent manufacturing industry,it has become an inevitable trend for robots to replace human beings in dexterous operation.6D object pose estimation technology is a basic task in robot operation,and it is also the premise that all operations can be operated flexibly and accurately.This task not only needs to locate the 3D position of the object,but also needs to estimate the orientation of the object.However,due to the complexity of illumination and occlusion in the real environment and the uncertainty of the shape and texture of the object to be operated,this task becomes more and more challenging.In recent years,due to the rapid development of deep learning in the field of computer vision,6D object pose estimation method based on convolutional neural network has become a hot research topic.In addition,due to the gradual maturity of depth camera,pose estimation based on RGBD image can make the whole model more adapt to the changes of environment.This paper studies the object 6D pose estimation network based on RGBD camera.The main work is as follows:(1)A novel multi branch network model based on RGBD is proposed for pose estimation of texture deficient objects.Based on the densefusion network architecture,the network adds a branch to extract the topological structure features of the object template in the feature extraction stage.The fusion of the topological features can effectively improve the role of depth information in object pose estimation.In addition,for the extreme imbalance between depth information and color information in the scene,a multi-scale feature fusion mechanism is designed to distribute the information of the two features,which effectively alleviates the influence of environmental factors such as strong illumination.Finally,a loss function based on the above information is designed to balance the losses of the two branches and improve the robustness of the network.Experiments on benchmark datasets show that the network effectively improves the accuracy of textureless object.(2)A network model that decouples rotation estimation and position estimation is proposed for 6D pose estimation of objects.Inspired by human habits,the network first estimates the rotation of the object relative to the camera,and then uses the rotation as a priori knowledge for accurate positioning.Aiming at the problem of different sources in point cloud registration,a source changing sampling mechanism is designed to make the object3D template and camera point cloud have similar topology,so as to carry out precise registration.Finally,a refined network based on rotation before translation is designed to adjust the object pose iteratively,so as to further improve the accuracy.The network can achieve the most advanced accuracy on both linemd dataset and YCB datasets dataset.And because the network is used to replace the traditional ICP algorithm for registration,it also has great advantages in reasoning speed.(3)A visual experiment platform is established to realize pose estimation in real scene.The platform uses Intel RealSense D415 as the sensor,establishes the instance segmentation dataset and pose estimation dataset in multiple scenes,and labels the dataset accurately using LabelMe and LabelFusion respectively.The effectiveness of the proposed algorithm is proved by experiments on the system.