Object Pose Estimation Using Point Pair Features

With the development of intelligent manufacturing,vision-guided robots are widely used in various manufacturing scenarios.Object pose estimation is a key technology for the robot vision system to perceive the surrounding environment,which can help the robot obtain the spatial position and orientation of the object in order to perform related operations.The problem of object pose estimation is defined as: given an object model,the algorithm estimates the spatial transformation of the object from the origin of coordinates to the actual scene.Object pose estimation methods are mainly divided into four categories: correspondence-based,template-based,voting-based,and regression-based.These methods have their own applicable scenarios.This paper focuses on the point pair features algorithm based on the voting method.The main research work is as follows:(1)A comparative study of the point pair features and their various variants is carried out.Firstly,using the interrelationships of the four dimensions of the point pair features,we designed variant forms of different dimensions;then,we introduced the color information encoded in different forms to the point pair features,and designed the variant form with color information.We designed the same experimental method and evaluation metric,and conducted comparative experiments on these two types of variants on five datasets.Experiment results show that the representation ability of the distance dimension is stronger than the angle dimension,and the color information helps to improve the feature recognition.(2)A method for object pose estimation using point pair features is proposed.The method is divided into two stages: offline modeling and online matching.It mainly includes 6 steps: preprocessing,feature extraction,feature matching,hypothesis generation,clustering,and post processing,and corresponding improvement strategies are introduced.We use the visible surface discrepancy function to measure the pose error,and use the mean recall as a performance indicator to conduct experiments on 6 datasets.We have analyzed in detail the improvement of sampling strategies,hash table unit access methods,bit arrays,number of pose hypotheses,ICP refinement and other improvement strategies on the performance of test results.Our method also achieves better performance scores than other types of methods.