Despite of many efforts by researches, real-time localization of mobile robot only using onboard sensor is still an open problem in mobile robotics. To contribute this research area, in this paper, a simple and effective method is proposed based on an adaptive estimator for estimating the position and linear velocity of a mobile robot in an unknown and unstructured environment using measurements of onboard inertial sensors and visual feedback of a catadioptric vision system. This novel method is developed on the basis of a new projection model of catadioptric vision, where the highly nonlinear projection model of catadioptric vision can be represented as a similar form to that of perspective vision. It is proved that this method leads to global exponential convergence of the estimation errors of the robot's position and linear velocity to zero. Experimental results validate the performance of the proposed method.