Object Detection and Localization Using Omnidirectional Vision in the RoboCup Environment

In this paper, a design and construction method for an omnidirectional vision system is described, including how to use it on autonomous soccer robots for object detection, localization and, also, collision avoidance in the middle size league of RoboCup. This vision system uses two mirrors, flat and hyperbolic. The flat mirror is used for detecting very close objects around the robot body and the hyperbolic one is used as a global viewing device to construct a world model for the soccer field. This world model contains information about the position and orientation of the robot itself and the position of other objects in a fixed coordinate system. In addition, a fast object detection method is introduced. It reduces the entire search space of an image into a small number of pixels, using a new idea that is called jump points. The objects are detected by examining the color of pixels overlapping these jump points and a few pixels in their neighborhood. Two fast and robust localization methods are introduced, using the angle of several fixed landmarks on the field and the perpendicular borderlines of the field. Borderline detection uses the clustering of candidate points and the Hough transform. In addition, the omnidirectional viewing system is combined with a front view that uses a plain CCD camera. This combination provided a total vision system solution that was tested in the RoboCup 2001 competitions in Seattle USA. Highly satisfactory results were obtained, both in object detection and localization in desired real-time speed.