 |
In the RoboCup Small-Size (F180) league, five robots on each team play soccer on a green field marked with white lines. The ball is orange and the robots, as well as the goals, are marked either yellow or blue. In addition to the yellow or blue team marker (a ping-pong ball centered on top of the robot), further markers are allowed, as long as they have different colors (refer to [6] for more details).
The robots are controlled by an external computer connected to a camera mounted above the field, such that the entire field is visible, as shown in Figure 1. The task of the vision system is to compute the positions and orientations of the robots, as well as the position of the ball. The behavior control software uses this information to operate the robots, relying on visual feedback. Since the robots and the ball move quickly and vision is usually the only input for behavior control, a fast and reliable computer vision system is essential for successful play. Further information about the overall system and the hierarchical reactive control of the F180 team FU-Fighters can be found in [1] and [2].
Appropriate coloring of the interesting objects partially simplifies the vision problem, but does not make it simple. There are several problems. First, the interesting objects are not always visible. The ball can be occluded, due to the central camera mount and vertical sides of robots. Also, the robots can be occluded by people that move them manually, e.g. to place them at their kickoff positions.
Inhomogeneous lighting is also problematic. There are shadows from the robots, the walls, and the referee, as well as highlights from the spot lighting. These conditions are far from a homogeneous diffuse lighting that would give the objects a constant appearance.
The vision problem is also complicated by the undefined background next to the field and the variability of the robots markers. All colors are allowed for the markers, as long as they are different from the reserved ones, and also the number and form of the markers changes from team to team. The vision system must be able to adapt quickly to the different markers of the opponents.
The non-linearity of the cameras wide angle optical system has also to be taken into account. The straight walls of the field appear to be convex in the captured image. One has also to correct for the height of the objects when mapping them to a 2D standard coordinate system.
The remainder of the paper is organized as follows. In the next section, we describe the robust tracking method, we developed. Then, we explain the non-linear coordinate transformation used. Section 4 presents improvements that we incorporated into the system prior the Melbourne competition: (a) the color map approach that models the appearance of colors dependent on the position; and (b) automatic identification of our robots.