logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-11102015-142553


Tipo di tesi
Tesi di laurea magistrale
Autore
MADEDDU, PAOLO
URN
etd-11102015-142553
Titolo
Development of an RGB-D image segmentation system for robotic grasping applications
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
INGEGNERIA INFORMATICA
Relatori
relatore Prof. Marcelloni, Francesco
correlatore Prof.ssa Lazzerini, Beatrice
correlatore Ing. Antonelli, Michela
Parole chiave
  • feature
  • object recognition
  • Point cloud
Data inizio appello
27/11/2015
Consultabilità
Completa
Riassunto
Image segmentation is the process of dividing an image into multiple parts following one or more criteria, with the purpose of simplifying its analysis.In robot grasping and object manipulation applications, the scene segmentation is the very first step. It consists in detecting which parts of an image represent objects and which one represents background. In this way, the robot can successively try to recognize each object separately and formulate a grasp hypothesis for each one of them.This work is part of PaCMan (Probabilistic and Compositional Representations for Object Manipulation), an European project with the goal of developing algorithms to allow robots to autonomously manipulate and grasp objects in domestic environments.The segmentation system proposed in this thesis starts following the classic approach of detecting horizontal planes in the scene; then, instead of directly removing them from the image, an average plane is computed and everything underneath it is filtered out. In this way, assuming there is only a dominant horizontal plane in the scene, the objects lying on it are isolated,if any. The objects found are then separated with an euclidean clustering algorithm. Also, some color-based clusters correction techniques are here presented. They come in useful in the event of wrong clustering, although they are valid only under some hypotheses. This thesis also includes a test environment used to measure the quality of the results. The tests are performed using two object recognition and pose estimation tools: the first makes use of global features, while the other one uses local features and it's meant to be used with a neural network. The design of such a neural network for classi cation and its application in combination with the aforementioned tools are part of this work.Both of the object recognition tools require an object database to search into; a database has then been built from the acquisitions of several objects, each one of them from diff erent angles.
File