• Object Recognition
• 3D Features
• Grasping
• Computer Vision
• Point Cloud
• Pose Estimation

Pose estimation serves as important tool for robot grasping applications, providing the necessary task-relevant information about the object that needs to be grasped.
Its use provide the robot with an estimation of the object geometry along with a full localization in 6 Degrees of Freedom of the object in space. These tools enable the
robot to manipulate the surrounding environment and grasp objects within, thus they are the first step towards the realization of autonomous mobile platform.

This thesis makes use of four global 3D features to gain multiple descriptions of the same object, then propose a combination of these descriptions, in effort to improve
the general performance and robustness of pose estimation procedure. We show how we can acquire meaningful data to build a database of features, so that an indexing and matching
procedure can take place, we'll then combine the responses into a list and use it to process our pose estimation.

A closer look on execution time will be kept, so that the pose estimation procedure could be run with real time constraints, if need be. The target robot, that will use this
procedure, needs to fast identify and localize objects within his environment, in order to competently manipulate them.

Along with data acquisition procedures, we propose some pre-processing pipelines to improve the general quality of our data and we show the benefits of good data pre-processing to
mitigate sensor imperfections and noise, that could be affecting the acquisitions.

The pose estimation procedure will be tested in numerous situations, including cluttered environment with both familiar and unfamiliar objects, treating both real and synthetic
data, to fully grasp its potentials and also limitations. The thesis is part of the European project Pacman: Probabilistic and Compositional Representations for Object
Manipulation, as a mean to establish bases to obtain robust grasping.