Research On Object Shape Recognition And Pose Estimation For Robotics Application

Object recognition is one of the most basic perception abilities of a robot. A robotcan grasp or track the object according to the perception information which provides theobject shape and pose information. As a high level visual feature, Shape is important tothe machine vision applications. By using the shape information, the robot can get moreinformation about the object, including the shape and the pose. How to represent andunderstand shapes in natural images is recognized as a fundamental problem. Till now, itis hard to get the ideal edge information from the natural image and it is also very hard tosegment the image. Most of the research is still focused on pure shape analysis. A mountof discoveries has been achieved, but few of them can be directly applied in practice.The research in this thesis is focused on vision based object recognition and poseestimation. The robot can perceive the object information from the environments. Byusing the vision based sensors, the features are extracted from the object. By utilizing themachine learning method, the object shape and pose information can be obtained. Thethesis mainly focuses on the shape representation, similarity measurement, shape based2D and 3D object recognition and 3D object pose estimation.Firstly, the research status on pure shape theory is analyzed. The global and localshape feature descriptors are discussed. The global shape feature descriptors are sensitiveto the noise and occlusion. And it is easy to appear the mismatch by using the local shapefeatures. Based on the intrinsic of the contour, a shape representation and shape matchingmethod is proposed, which combines local and global shape information. The shapes arealigned by using the global shape descriptor. Several local shape descriptors are combinedto describe the local shape. The shape matching similarity is computed by using all thefeature descriptors. The proposed shape matching method has good performance on com-putation e?ciency. And It is applied to shape matching to achieve the good performanceon MPEG7 and KIMIA databases.Secondly, the challenges and di?culties in the ?eld of natural image object recog-nition are discussed. It is di?cult to realize the natural image object recognition forseveral reasons. Most of the shape cannot be represented by the parametric equations,and the edges extracted from the natural images are sensitive to the noises and viewpointchanges. For those reasons, an object recognition method based on the shape fragments is proposed. The descriptor of the shape fragment is invariant to scale, noise and rota-tion. By using the proposed shape fragment descriptor, the proposed method which underthe top-town object recognition frame can achieve good results on ETHZ data set andacquired real images.Then, the 3D object modeling methods are analyzed. It is hard to model the sameclass 3D object. And it is also very hard to model the ?exible 3D objects and complex3D objects. Under di?erent viewpoints, the appearance of the object is di?erent. The3D object can be modelled by multi-view images. Based on the research of multi-viewimage features and multi-view geometry, a 3D object recognition method is proposed.The proposed method is based on the multi-view mixture model. A learning method isapplied to the proposed method. The mixture model is composed by the point model andshape fragment model. Di?erent viewpoints are connected by the multi-view geometry.The 3D object model is composed by the views, morphed views and the multi-view ge-ometry. The proposed algorithm can achieve good results on Stanford 3D object data setand acquired videos.Finally, 3D object pose estimation problem is analyzed and it can be in?uenced bymatching precision, noise, optimization algorithm and so on. 3D object pose estimationmethod is proposed by using the shape fragments matching. The object pose can be esti-mated by a single image. The high precision pose can be achieved by the pose estimationequation which satis?es the Sylvester's equation. Shape information is a research foun-dation in this thesis. The research includes several key problems in the ?eld of objectrecognition and object pose estimation. The proposed algorithms are validated by sever-al experiments. The theory and its applications of the robot vision can bene?t from thetheory, models and algorithms proposed in this thesis.