Intelligent Global Vision for Teams of Mobile Robots

摘要

Mobile Robots, Perception & Navigation vision will also mean that there will be significant redundancy in processing across a team of robots in many applications as well. Local vision may also be undesirable in applications where a large number of very simple robots may be able to do the job of a few complex robots, in environments where shared vision is amenable (that is, where a unique perspective for each individual is unnecessary), and in educational environments where it is desirable to separate the problems of computer vision from the rest of robotics. In these domains, the second form of vision, global vision, is often preferred. Global vision provides a single third-party view to all members of a robot team, analogous to the view of a commentator in a soccer game. Global vision shares many of the problems associated with local vision. Objects of interest must be identified and tracked, which requires dealing with changes in appearance due to lighting variation and perspective. Since objects may not be identifiable in every frame, tracking objects across different frames is often necessary even if the objects are not mobile. The problem of identifying objects that are juxtaposed being viewed as one larger object rather than several distinct objects, and other problems related to the placement and motion of objects in the environment, are also common. In domains such as robotic soccer, where pragmatic real-time global vision is large part of the application, many of the more difficult problems associated with global vision have been dealt with through the introduction of artificial assumptions that greatly simplify the situation. The cost of such assumptions is that of generality: such systems can only operate where the assumptions they rely upon can be made. For example, global vision systems for robotic soccer (e.g. If a camera cannot be placed perfectly overhead, these systems cannot be used. Such systems also typically recognize individuals by arrangements of coloured patches, where the colours (for the patches and other items such as the ball) must be pre-defined, necessitating constant camera recalibration as lighting changes. Such systems can thus only operate in environments where lighting remains relatively consistent. While such systems will always be applicable in narrow domains where these assumptions can be made to hold, the generality lost in continuing to adhere to these assumptions serves to limit the applicability of these approaches to harder problems. Moreover, these systems bear little resemblance to human vision: children playing with remote-controlled devices, for example, do not have to climb to the ceiling and look down from overhead. Similarly, human vision does not require significant restrictions lighting consistency, nor any specialized markings on objects to be tracked. In order to advance the state of the art in robotics and artificial intelligence, we must begin to make such systems more generally intelligent. The most obvious first steps in this direction are considering the assumptions necessary to make a global vision system operate, and then to find ways of removing these. O u r a p p r o a c h t o r e a l t i m e c o m p u t e r v i s i o n a r i s e s f r o m a d e s i r e t o r e m o v e t h e s e assumptions and produce a more intelligent approach to global vision for teams of robots, not only for the sake of technological advancement, but from a pragmatic standpoint as well. For example, a system that does not assume that a camera has a perfect overhead mount is not only more generally useful, but requires less set-up time in that a perfect overhead mount does not need to be made. Similarly, an approach that can function in a wide range of lighting conditions saves the time and expense of providing specialized