Back to search

Democritus University of Thrace (DUTH)

New pattern recognition techniques in machine vision systems

Abstract

dc:description

As robots’ working environments become more and more complex, new pattern recognition challenges emerge. Apart from the conventional question ”What is the object?", nowadays, scholars should answer to other appealing queries such as "Where is it?" or "How to manipulate it?". State of the art vision methods consist of advanced recognition subroutines that are fed into machine learning algorithms so as to reach assumptions, regarding the current world state, akin to the ones of a human mind. One of the most challenging tasks in machine vision is the estimation of the 3D pose of an object, due to its practical implication and its augmentative importance in scene interpretation processes. A plethora of diverse applications impose upon the accurate assessment of objects’ geometrical configuration, relative to a given coordinate system, in order to achieve their goals. In robotics, autonomous object manipulation can efficiently be accomplished only in cases where the full pose (six Degrees of Freedom (DoF)) of the testing target is known. Despite the affluent research endeavors and the achievements reached so far, an advanced vision system characterized by low complexity and sufficient generalization capabilities has yet to be built. This PhD thesis is inspired by the remarkable skills of humans in the particular task of estimating the relative pose of objects given an initial hypothesis. Regardless the lighting conditions and common visual disturbances, such as partial occlusions, humans excel in interpreting the 3D geometrical configuration of arbitrary placed objects. Among all operations realized by human beings, the majority is directly related to object manipulation. On the other hand, the realization of target manipulation by a robotic arm puts strict constraints regarding the exact speed of execution and the "nature" of the testing objects. Contemporary solutions to this problem require high storage capabilities whilst failing to generalize to unknown testing targets, i.e. objects that are not included into the training set. The main objective of this PhD thesis is the development of innovative vision algorithms tackling the 3D object pose estimation problem and that are to be realized on robotic platforms. It is apparent that, the designed architectures should initially, deal with the position estimation problem -i.e. mainly the depth estimation one- and gradually solve for the remainder three orientation DoFs. The relevant literature was surveyed to reveal the current state of the art and to highlight open research issues that are to be addressed. Afterwards, novel manifold modeling approaches, resulting in low dimensional and high discriminative feature vectors, were designed and implemented. Lastly, the developed algorithms were adopted by robotic platforms to carry out autonomous object grasping. In the first research endeavor of this thesis, the construction of a simple and easy-to-build framework for location assignment of an object in a scene, is investigated. The developed method is based on the observation that features extracted from any two-part (detector-descriptor) algorithm correspond to spots on the object’s surface and their center of mass is related to the one of the objects. Thus, by extracting these features at known positions of the sought object, one can estimate its distance from the camera. Comparing to the contemporary solutions for depth estimation, the employed approach is computationally inexpressive whilst requiring only a single sensor opposed to stereo vision architectures. Moreover, a neural network-based framework that is able to calculate the three rotational DoFs of any object is presented. The network is trained with numerous targets contained in several available datasets with the training process been guided by a fuzzy extraction of the centers of the abstracted features through the Fuzzy c-means algorithm. The 3D pose of an object in a training instance is considered as the distances of the fuzzy centers from one particular cluster. The proposed method involves a new input-output mapping that reduces the dimensionality of the input vectors with good performance. The contribution of this study entails the formalization of this new input-output method that outperforms the conventional dimensionality reduction techniques widely used in image processing applications. Following the intuition that (a) one object viewed under varying perspectives lays on a well-projected subspace and (b) different objects captured under similar viewpoints share identical poses, a sophisticated framework capable of both recognizing objects and estimating their pose in the 3D working space was coined. It encompasses a manifold modeling procedure that depends on the attributes of mutual information and a constellation-based structure, respectively. The employed recognition module incorporates a modified approach of a known dimensionality reduction technique that constructs a similarity matrix based on mutual information among objects and then seek a low-dimensional representation that preserves the local structure of the objects in the initial high-dimensional space. Once the data are projected into the sub-space a Support Vector Machine (SVM) classifier provides accurate recognition. The 3D pose estimation module is based on a manifold modeling algorithm that constructs feature vectors of low dimensionality and high discriminative capabilities. Then a Sparse Grasping Manifolds (SGM) method is proposed in this thesis, which aims at solving the problem of manipulation of unknown objects by unifying 3D pose manifolds and grasping points into a cohesive framework for robot grasping. Unlike contemporary systems that crave extensive supervision and large repositories of images of objects, meticulous emphasis is given on providing a ground solution with large generalization capacities founded on unsupervised learning. Additionally, the visual data available are processed in a way that their projection onto the corresponding subspaces is sparse, compact and highly representative. Grasping manifolds depend on a novel bunch-based architecture, here introduced for the first time that, unlike previous works, bypasses the part selection process using unsupervised clustering, while by extracting local patches encapsulates both appearance and geometrical attributes of the objects. Contrary to earlier works the presented method offers higher generalization capacities mainly due to the efficient learning of the employed function that is based on a large a priori training set containing numerous examples of real and artificial data. Moreover, through the minimization of the l₁ norm, the presented approach builds sparse and compact manifolds that are highly representative, tolerant to common imaging disturbances (e.g. noise) and possess superior discrimination capabilities. In the last study of this thesis, the goal was to design a unified architecture for autonomous manipulation of unknown objects, which is capable of answering and addressing the constraints of all the following questions: "What is the object?", "Where is it?" and "How to grasp it?". The recognition problem is tackled from a shape-based perspective, whilst obtaining accurate detection decisions via a Bag-of-Visual Words classification scheme. Grasping points are found through an ontology-based knowledge acquisition where recognized objects inherit the grasping points assigned to the respective class. The presented ontologies include: a) object-class related data, b) pose-manifolds assigned to each instance of the object-class conceptual model and c) information about the grasping points of every trained instance. This work represents the first integrated research endeavor in concept ontologization focusing on the liaison between image understanding algorithms and the corresponding motor commands in the particular task of unknown object grasping.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2013

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kouskouridas, Rigas
  • Κουσκουρίδας, Ρήγας

Subjects

dc:subject × 16

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/29283

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Kouskouridas, Rigas; Κουσκουρίδας, Ρήγας. New pattern recognition techniques in machine vision systems. Democritus University of Thrace (DUTH), 2013. http://hdl.handle.net/10442/hedi/29283