Towards Bootstrapping Robotic Perception of Indoor Environments

Towards Bootstrapping Robotic Perception of Indoor Environments

Alagarsamybalasubramanian, A.C.

Jonker, P.P. (mentor)
Rudinac, M. (mentor)

Mechanical, Maritime and Materials Engineering

BioMechanical Engineering

BMD

2012-08-28

Robots have been popular amongst us as Science fiction characters for a few decades now. The inability of the robots to robustly perceive and respond to the real world has been confining them to the laboratories for a long time. This can be attributed to the dynamic nature of the everyday environments where the prelearnt knowledge alone is not sufficient. The robots can be developed to work autonomously in these situations when they can obtain and update the knowledge of their surrounding on their own without external intervention. This process is termed as bootstrapping and it involves perceiving various aspects of the environment like faces, objects, sound, etc. This is a very elaborate task given the current level of sophistication. It is important for an intelligent robot to be able to comprehend unknown objects in the scene and this thesis focuses on handling unknown objects. A segmentation based on the 3 dimensional sensory data of the visual scene is implemented. The large planar structures in the scene are identified and the rest of the data is clustered to locate the probable objects in the scene. The appearance of the objects are chromatically and spatially described and are the basis on which they are recognized from the pre-learnt object models. A generic grasping technique based on the visual tracking feedback has been proposed and implemented to grasp any different object from various common locations.

robots
perception
object recognition

http://resolver.tudelft.nl/uuid:ae201f3b-6470-458e-a988-f077484a6639

2013-08-28

Student theses

master thesis

(c) 2012 Alagarsamybalasubramanian, A.C.