Active Vision and Robotics

July 2005, I started my PhD project. The main themes of my research are active vision and 3D object recognition. Active vision refers to the active capabilities of a system to steer its visual input. Doing so, a system can actively change its viewpoint to get more and better information than a passive viewing system would be able to.

My research consists of two projects. In one project, we look at human eye movements and try to predict the eye fixations. In the other project, we investigate how a robotic system can learn to recognize 3D objects in the real world. I will briefly outline both projects.

Project 1: Predicting Human Eye Movements


Eye movements
on a scene.
When we humans view a scene or an object, our eyes perform a series of eye movements. The eyes shortly fixate parts of the object on the fovea, which is the high-resolution part of the retina. By doing so, the computational load is kept low, while high-resolution vision is obtained.

The eyes do not simply scan the environment randomly. On the contrary, only the interesting parts of the environment are paid attention to. In this project, we investigate whether we can predict these eye movements with image processing techniques. To investigate this, we compare eye tracker data, obtained from experiments with human participants, with region-of-interest detectors.

Project 2: Learning 3D Objects Through Active Vision


A 15 month old child exploring an
object
In the real world, the object constancy problem plays a huge role. Object constancy refers to the problem that objects can look very different from different viewpoints. Features like shape, color and texture change, but still it remains one (constant) object. To deal with this, we humans (and other animals) explore objects by turning it around and look at it from different orientations. When we cannot recognize an object, we behave similar, in order to get more and perhaps better information about the object. In this second project, we try to copy this behavior, in order to let a robot learn to recognize objects in the 3D world.

Combining the Projects: Using Eye Movements for Object Recognition

When a robot views a scene or an object, it also has to deal with an enormous amount of visual information, similar to human visual system. In order to efficiently represent the scene, the robot also has to filter out the relevant and interesting information. Therefore, the results of the first study will be applicable in the second study. The region-of-interest detectors resulting from the first study will be used to improve object recognition in the second study. In this way, we hope to learn from solutions found in nature to solve problems in artificial systems.