Exploitable results
Robots doing domestic chores have long been the holy grail. These machines may be just a decade away, believes the leader of a European project studying the nuts and bolts of computer vision. Humans can easily recognise and categorise things or events. But computers, even after four decades of experimentation, still struggle to do the same. While researchers from various disciplines have learnt how to signal process, recognise and interpret images, there has been little collaboration among the communities.
The Information Society Technologies programme-funded project COGVIS has spent two years advancing computer vision. "Our consortium has researchers who study human and traditional computer vision, signal processing, learning and artificial intelligence," says Henrik Christensen, coordinator of the initiative from Stockholm. "Now is the time to integrate all these disciplines and build systems that can recognise things and learn, and then know what to do."
Computer vision is successfully used for quality-control processes in specific fields such as face recognition and basic visual surveillance. But cognitive vision is multi-faceted, complicating this domain for computers. "Early work on vision assumed that you only needed to generate a 3D model of the world, then interpret the geometry of that model and generate a symbolic world," says Christensen, adding that it is unfortunately impossible to generate robust 3D models.
Recent work on vision has broken down into research guided by artificial intelligence and geometry - but with little integration of the two approaches. Vision has also typically been considered in isolation. Yet it needs to be considered in the context of its task. "You should know what vision is going to be used for," explains Christensen.
Know what you are doing
He asks us to imagine a table with various objects on it, including several cups. A computer vision system could compute a complete depth map of the scene and from that map construct a complete geometry model. The process resembles computer-aided design. Next, the system could try to locate and grasp a cup or only search for the cup that it wants to pick up. "If a machine or person knows what it's doing, the work is much simpler," he adds. "This fact was only recently recognised."
The project aims to provide the methods and techniques needed for construction of vision systems. Such systems will perform task-oriented categorisation and recognition of objects and events, in the context of an embodied agent. A robot, for instance, could interpret the actions of humans and interact with its environment - perhaps fetching and delivering objects in a home.
There are many challenges. For example, in this illustration, computer vision would find it difficult to distinguish between pictures and real objects. The chair in the back room appears to be the same size as the object on the table. Only the context - the space and position of these chairs in the scene - reveals one is a toy and the other a real chair. "No single technique, using basic clues, would recognise all the chairs here," admits Christensen. "Today we lack techniques that allow us to correctly classify all objects in this image."
The project has divided its research into four areas: recognition and categorisation of objects, structures and events; reasoning about and interpretation of scenes and events; learning and adaptation; and control and integration. The emphasis is on studying these systems together, in realistic settings.
Although COGVIS focuses on basic research, it has been building robots to test out ideas. The project's Hamburg partners have put together a machine that interprets what is going on at a dinner table. "It can encode social and cultural knowledge," says the coordinator, "using, for example, the location of knives and forks to identify plates." The team in Sweden is building a robot that can clear dinner tables.
All these efforts require a combination of recognition strategies, including those based on appearance and geometry. But they could also be extended to an object's colour and texture as well as non-visual characteristics.
The project is also studying the difference between recognition and characterisation. "This will enable us to build systems that are useful across a variety of domains, such as recognising different kinds of plates," says Christensen.
A demographic driver
"My motivation for this work is Europe's demographic profile," he says. "Over the next 20 years, there will be a 50 per cent increase in the number of people of retirement age. And there will be fewer workers to support them." Christensen is determined to find ways of increasing working efficiency for all, as well as helping the elderly to remain autonomous.
Machines are an obvious solution. "We want to build systems that can work in any home and work within any environment," he adds. That will require further research on priorities such as categorisation, learning, artificial intelligence and representation - and then combining them all.
He admits that such versatile robots, especially those able to interact with humans, cannot be built today. But they could be here within just 10 to 15 years, autonomously carrying out tasks such as managing logistics systems in factories, vacuum cleaning, and assisting elderly people and the handicapped with meal preparation and basic cleaning. Several Nordic handicap associations are showing interest in the latter aspect of the project.
At the project's June meeting, held in Leeds, England, various teams presented the results of their work. Achievements include the development of a model based on how humans recognise things; a more efficient model for learning of objects; a new model for interpreting a dinner set-up; and a robot called Baby Bot. The robot, developed in Italy, is studying arm/hand movement and grasping skills.
Autonomous systems have many potential applications, notes Christensen. They could act completely on their own in unfriendly or inaccessible environments. Others could be used as dextrous assistants to humans doing dull or difficult tasks.
One of this project's results - integration of interpretation and modelling of human interaction - could be used in video streaming. For example, during televised sports programmes, video recorders could be told to find and display a specific goal or action. Other future applications include better quality control in factories, automatic detection of unsociable behaviour, and instruction by demonstration - where a robot would learn by watching a human doing a task rather than being programmed.
Source: Based on information from COGVIS
Promoted by the IST Results Service
Wyszukiwanie danych OpenAIRE...
Podczas wyszukiwania danych OpenAIRE wystąpił błąd
Brak wyników