MCCOOP

For some 30 years, scientists have aspired in developing a computer with intelligent visual skills, that is able to recognise and interpret objects or events. Early efforts, which involved either mathematical modeling or attempts to imitate the processing of visual data by the brain, had limited success. Now the MCCOOP project is bringing together the two approaches. Its coordinator, the Fraunhofer Institute for Applied Information Technology (FIT), is experienced in biologically motivated computer vision and the provision of advanced visual technologies for industry, like algorithms for autonomous robot navigation. The other project partners (Computational Neuroscience Research Group, from Katholieke Universiteit Leuven, Belgium; PSPC-Group from University of Genoa, Italy) supply expertise in neurophysiology and biophysics. Edges and motion

Visual nerve cells (neurons) are responsive to various cues, such as motion, texture, colour and disparity, allowing cell groupings to detect, segregate and code a variety of objects in complex surroundings. MCCOOP's aim is to use the existing neurophysiological studies to relate the anatomy and physiology of visual neurons to the way specialized assemblies of neurons detect, segregate, and code objects. Translating this knowledge into computer vision, the project hopes to develop a mechanism for detecting the edges of objects based on a combination of visual cues - in addition to orientation - to enhance how the contours of moving objects are extracted.

The project will also work on detecting independently moving objects, aiming to distinguish between different trajectories of each object and the self-motion of the observer by linking information on motion and binocular disparity (depth).

While many applications of computer vision will benefit from these studies, their application to videos of the surroundings viewed by a car driver will offer very substantial improvements on the computer visualisations currently available. The great advantage of MCCOOP is that it can estimate multiple features, such as contour, motion, heading and stereo, in order to provide far more accurate detection of cars, pedestrians, etc. The use of an individual sensor to determine one aspect of the visual environment - for example, object edges - needs weighting to compensate for the sensor type and its limitations. The MCCOOP method generates a single integrated signal ('early joint code') comprising the interaction of multiple visual cues.

Industrial applications

MCCOOP is a highly innovative project which involves a high technical risk as numerous previous attempts have failed to develop a model for computer vision that is generally applicable. The project is seeking a breakthrough in solving complex image-processing problems: a neuron-based system capable of flexibility within a wide range of conditions, using co-operative processing. The same principles can then be applied to the development of a new generation of image-processing algorithms. If successful, the project's impact will be vast, as once it is sufficiently reactive and flexible, computer vision will be very suitable for a large number of industrial applications, such as quality control and assembly, monitoring traffic safety, and building security. It could even provide sight reinforcement for the visually impaired.

In the area chosen to test the system developed within the project, a computer vision system for analyzing the scenes confronting drivers would increase their safety and offer important benefits in the development of new car models. In use, it could interpret potentially dangerous information like a car crossing the direction of intended travel, and trigger suggested steering or braking signals. Several major car manufacturers are already engaged in car-safety programmes which could benefit from the MCCOOP results, and consultations with the car industry will be organized along with more general scientific publication and dissemination of the project's results.

Contact

Dr. Marina Kolesnik

Phone: +49 (0) 2241/14-3421

marina.kolesnik@fit.fraunhofer.de

Official project page

http://www.mccoop.de (probably dead by now).

Partners

• • Fraunhofer Institute for Applied Information Technology (FIT), Sankt Augustin, Germany (coordinator);

  • Computational Neuroscience Research Group, Katholieke Universiteit Leuven, Belgium;

  • PSPC-Group, University of Genoa, Italy.