Object Operation Oriented Pose Estimation Of House-Hold Objects Based On Improved Point Pair Features

In family environment with a large number of texture-less objects,whether the robot can accurately estimate the pose of the target object and operate the object rea-sonably according to the task requirements are the key to the robot's operation task,and these are important parts of the robot's ability to provide intelligent services.Current pose estimation methods are susceptible to similar surfaces in the environment,irrele-vant scene information will also reduce the speed of calculation,and the occlusion of objects in the home will also affect the accuracy of pose estimation.Current pose esti-mation methods only provide the robot with the pose of objects,and lacks semantic information related to the operation of the object.Therefore,the relationship between object and operation task cannot be established,and the robot cannot complete the task of object operation more intelligently.In the pose estimation methods of weakly tex-tured objects,PPF is a robust feature used to describe the surface of a 3D object,which can obtain higher pose estimation accuracy;In addition,by designing the general oper-ation attributes of objects to provide robots with the knowledge necessary to complete the manipulation tasks of objects.This paper draws on image segmentation technology to study and improve the pose estimation method of PPF.We have designed semantic descriptions related to the operation of objects and proposed a method of generating operational knowledge.These methods can improve the accuracy and speed of pose estimation,and at the same time improve the intelligence level of robot manipulation.The main tasks completed are as follows:First,a set of operation-related object attributes are designed for obj ect functions,which are used as object operation knowledge to be provided to the robot to plan the operation of the object after pose estimation is completed.A commercial depth camera is used to obtain a three-dimensional model of the object by using the obj ect reconstruc-tion method,and object is divided into multiple parts,each of which has a different function or shape.The operation types of objects and components are divided into four categories,and the operation attributes of each category are specified.The operation attributes of each operation mode can describe the operation methods to be used to complete the corresponding functions of the objects.Use existing point cloud annota-tion tools to label parts of an object and set operational properties for the entire object and each part.Second,the point pair features method is used to estimate the six-dimensional pose of the object in the extracted target scene,which can reduce the interference of irrele-vant information in the scene and reduce the amount of information,thereby improving the calculation speed.The K-means clustering method is used to divide the color of the object into multiple clusters,and the color characteristics of the item are obtained through modification.The scene region where the target object may exist is extracted by using object color features.And point pair features are calculated on the target scene region.Multiple candidate poses obtained by voting for the overall features of object.The higher the confidence of the pose,the more likely it is the correct pose.Thirdly,the voting method of point pair feature is improved to reduce the interfer-ence of similar surfaces in the scene,and reduce the impact of occlusions.The cumula-tive matrix is used to count the number of pose votes.Each reference point in the scene will obtain a pose with highest vote.The highest vote is modified by using the votes of the adjacent poses.Proposed pose estimation methods are verified on the public occlu-sion dataset.The experimental results show that the proposed region extraction and vote correction methods can improve the accuracy and speed of pose estimation.Finally,we propose an operation task-oriented operation knowledge generation method to verify the rationality of 3D model operation attributes.According to the pose of object,combined with the operation task to be completed and the operation attribute of the object,the positional relationship between the object and the environment is an-alyzed to determine whether the object satisfies the status required to complete the op-eration task and whether the operation part of the object is available.If there is a con-dition that is not satisfied,the pose of object needs to be adjusted,and the operation knowledge of the object is output when the condition is completely satisfied.The ef-fectiveness of the proposed method is verified by the operation knowledge generation experiment when performing "move" and "catch water " tasks on the cup in different pose.