By Noah Shachtman
July 16, 2008 12:01:00 PM
Categories: Drones
Courtesy Of Wired.com
Machines and robots still have a hard time figuring out what they're seeing. Which is why the Army wants 'em to start sorting through images more like human beings.
The Army recently put out a call for a "psychologically inspired object recognition system... Such a system would be extremely beneficial for robotic control/intelligence and would allow for an exponential expansion of robotic capabilities and intelligence."
If it all works out as planned, the Army thinks the new robo-vision 'ware could be used in "robotic security systems, autonomous factory systems and robotic health care systems," as well.
"Recognizing and identifying an object from a video input turns out to be a very difficult problem. The problem stems from the fact that a single object can be viewed from an infinite number of ways," the Army sighs.
Sometimes, people use "template based algorithms" -- like matching an object to a database. Sometimes, they look for particular features. In other situations, they watch for geometric icons, or geons. "These three algorithms are used in conjunction with a fourth algorithm, a contextual cueing algorithm, which limits the overall search space. Finally, human spatial memory is able to mentally rotate objects in order to match the object to different representations."
So the idea is to combine all the methods, to produce a more reliable robotic recognition system. By the end of the second phase of the program, the Army wants a working prototype that can ID objects -- regardless of scale, regardless of rotation -- at a rate of 95 percent or better.
July 16, 2008 12:01:00 PM
Categories: Drones
Courtesy Of Wired.com
Machines and robots still have a hard time figuring out what they're seeing. Which is why the Army wants 'em to start sorting through images more like human beings.
The Army recently put out a call for a "psychologically inspired object recognition system... Such a system would be extremely beneficial for robotic control/intelligence and would allow for an exponential expansion of robotic capabilities and intelligence."
If it all works out as planned, the Army thinks the new robo-vision 'ware could be used in "robotic security systems, autonomous factory systems and robotic health care systems," as well.
"Recognizing and identifying an object from a video input turns out to be a very difficult problem. The problem stems from the fact that a single object can be viewed from an infinite number of ways," the Army sighs.
By rotating, obscuring, or scaling a single object, one can create multiple representations of an object - which makes the problem of matching the object to a database of objects very difficult.One way around all this may be to figure out "how human beings recognize objects in the real world and duplicate that functionality in a series of algorithms." But here's the catch: "Recent research has indicated that humans use not one algorithm, but multiple algorithms for the task of object recognition - depending on the object being recognized and the situation at hand."
The problem expands exponentially when objects that need to be identified have never been viewed before. Combine these limitations with the wide variety of objects which might be identified, and the problem becomes intractable.
Sometimes, people use "template based algorithms" -- like matching an object to a database. Sometimes, they look for particular features. In other situations, they watch for geometric icons, or geons. "These three algorithms are used in conjunction with a fourth algorithm, a contextual cueing algorithm, which limits the overall search space. Finally, human spatial memory is able to mentally rotate objects in order to match the object to different representations."
So the idea is to combine all the methods, to produce a more reliable robotic recognition system. By the end of the second phase of the program, the Army wants a working prototype that can ID objects -- regardless of scale, regardless of rotation -- at a rate of 95 percent or better.
No comments:
Post a Comment