This paper studies the problem of moving target tracking in 2D space by omnidirectional mobile robots using stereo vision. For a given moving target, by extracting its feature points using vision, the relative motion, including both position and velocity errors between the target and the robot can be estimated. Unlike the case of nonholonomic robot tracking whose path planning is a difficult task, in our vision based tracking of omnidirectional robots, the tracking path is directly specified by the estimated tracking errors. Given the dynamics of robots, a PD control is proposed which ensures the asymptotic stability of the tracking error if the target is moving with constant linear and rotational velocities, or the bounded-input-bounded-output (BIBO) stability if the target is moving with varying but bounded rotational and linear accelerations. Simulation results show the efficiency of the algorithm.