SPREKER

Maja Rudinac, promovenda, TU Delft

TITEL

Visual categorization of unknown objects for mobile robotics applications

ABSTRACT

One of the main challenges of modern robotics is designing robots able to explore novel environments and learn from the surrounding. Focus of this presentation is our framework for active scene exploration and learning of objects that is suitable for robotic applications in the dynamic environments.

First step in categorizing and learning novel objects is detecting their location for which we designed an object localization module. This module is inspired by human visual system and is performing segmentation of salient (interesting) objects based on their distinctiveness from the rest of the scene.

Once the objects of interest are localized, attention of the system is focused on them and object inspection module is deployed.

The object inspection module has two parts. The first part can recognize and learn specific objects from multiple viewpoints using our robust descriptor based on the combination of color, texture and shape information. The second part of the module is dedicated to object categorization. In order to adapt to new environments, the system must be able to transfer knowledge learned from previous experience. Therefore, the system must be able to classify objects in meaningful categories and link all knowledge about the objects’ properties to the categories they belong to. Since all similar environments are characterized by a limited number of object categories, (such as e.g. ‘chairs’, ‘computers’ or ‘tables’ in an office environment), if system learns these categories, it could more easily adapt to an unknown environment.

To perform categorization, we offline construct a database of objects belonging to different categories and represent each category by a descriptor that describes average visual representation of all objects in that category (in the form of a visual codebook). In the online step, the object is inspected from several accessible viewpoints and descriptors are extracted and matched with the database to detect the category to which a particular object belongs. If the object is not found in the database, its category can still be recognized and the object can be learned. Experiments performed in both experimental and real-world environments confirm viability of our approach.

BIOGRAFIE

Maja Rudinac is a PhD student in the Biorobotics Lab at the Delft University of Technology. Her research interests cover a broad field of computer and robot vision as well as the applications of machine learning in robotics. The main idea of her thesis is the design of a robot vision system capable to detect and recognize specific objects and object classes in the indoor environment and learn their visual properties by interacting with them and inquiring a human supervisor. She is a member of the Esi project Falcon, which concentrates on a new generation of distribution centres and warehouses with a maximum degree of automation. Her part includes designing the vision and pattern recognition system for the automated picking and stacking of objects.