by SK » Wed Dec 29, 2010 10:45 am
by SK
Wed Dec 29, 2010 10:45 am
Most RoboCup teams (in the humanoid and standard platform league, that is) use largely custom image processing approaches, as there is very limited processing power available on the robots and it thus makes sense to use all available knowledge and possible optimizations for the given scenario.
In both leagues, elements of the playing field and robots are currently color-coded and this fact is leveraged in image processing systems. Both B-Human and the Dribblers (champions 2009 and 2010 in their respective leagues) use a so called "scanline" approach, where the image gets sampled using scanlines that are more coarse than the image resolution. For all sample points, the pixel gets assigned a color class using a look-up table and based on this, with further processing, objects can be detected.
Of course, the playing field will not stay color coded forever and the rules are moving a small step towards a real soccer scenario every year, so all that will change over the years and with more processing power available in smaller packages we'll likely also see more use of OpenCV in RoboCup humanoids.
All that of course does not mean that OpenCV is "bad", it only means that with the severe constraints of the RoboCup humanoid scenario, custom solutions are currently more successful.
Most RoboCup teams (in the humanoid and standard platform league, that is) use largely custom image processing approaches, as there is very limited processing power available on the robots and it thus makes sense to use all available knowledge and possible optimizations for the given scenario.
In both leagues, elements of the playing field and robots are currently color-coded and this fact is leveraged in image processing systems. Both B-Human and the Dribblers (champions 2009 and 2010 in their respective leagues) use a so called "scanline" approach, where the image gets sampled using scanlines that are more coarse than the image resolution. For all sample points, the pixel gets assigned a color class using a look-up table and based on this, with further processing, objects can be detected.
Of course, the playing field will not stay color coded forever and the rules are moving a small step towards a real soccer scenario every year, so all that will change over the years and with more processing power available in smaller packages we'll likely also see more use of OpenCV in RoboCup humanoids.
All that of course does not mean that OpenCV is "bad", it only means that with the severe constraints of the RoboCup humanoid scenario, custom solutions are currently more successful.