In this decade, a number of various robot contests represented by "RoboCup" have been held in various parts of the world. In some of these contests, such as RoboCup middle-size robot league, fully automated mobile robots are used to compete for their mobility and autonomy. These robots have to plan optimum moving strategy automatically without any human intervention after the game starts. One of the fundamental and crucial faculties for these autonomous robots is the ability of accurate positioning of enemy and themselves.<BR>In this paper, a robust positioning system is proposed for soccer robots equipped with an omni-directional camera and wheels. The robot identifies its own position by observing some landmarks placed around the field through an omni-directional camera, and by counting rotation of wheels. To achieve the robust positioning, the proposed method utilizes (1) LMedS method for making correct correspondence between observed directions and landmarks, (2) elimination of miss-correspondence of landmarks using a median filter, (3) position calculation considering the properties of measurement error distributions using nonlinear maximum likelihood method, and (4) data fusion of positions obtained by observing landmarks and dead reckoning using kalman filter.