Using pattern matching on a flexible, horizon-aligned grid for robotic vision

Image processing in the robotic soccer domain meets real time constraints that must be met with optimizations of the task. We are proposing a method to find objects using a non-uniformly spaced, optimally constructed grid. We are capitalizing on the fact that the apparent size of ground-based objects is related to their visible distance to the horizon. In a three-stage process, a pattern matching is performed along the grid-positions, object candidates are identified by specialized modules, and finally, the consistency of the result is improved. We describe an example identificator module in more detail.