This thesis describes the design and implementation of a global vision system for robot soccer. This system was designed for use with the University of Queensland’s team, the RoboRoos, in the 1999 Robot Soccer World Cup. Although functional, previous attempts at solving the robot soccer vision problem have lacked robustness, being overly sensitive to changes in environmental lighting conditions. Many of these attempts also made use of a tracking based solution, which left them prone to nonrecoverable fault conditions caused by the loss of a robot’s identity.
The global vision system works in two parts. The first of these sections is responsible for segmenting the robots and the ball from the background field. This segmentation is achieved by first determining which parts of the RGB colour space correspond to field green, and then performing a thresholding algorithm based on this information. Noise is then removed from the thresholded image, and objects are formed using a standard region-growing algorithm. Once the robots and the ball have been isolated from the background field, information can be extracted about them. The object recognition section of the system uses template matching and a unique method of encoding information in markings on the top of each robot to determine the orientations and identities of each of the RoboRoos.
The final vision system was able to provide all information required, and to operate in a manner that ensured the reliability, accuracy and speed of the information. The system also proved robust to changes in lighting and to fault conditions. This assessment was confirmed by the strong performance of the global vision system in the RoboCup ’99 competition.