We present a novel practical method for self-calibrating a camera which may move freely in space while changing it internal parameters by zooming. We show that point correspondences between a pair of images, and the fundamental matrix computed from these point correspondences, are sufficient to recover the internal parameters of a camera. Unlike other methods, no calibration object with known 3-D shape is required and no limitation are put on the unknown motion, as long as the camera is projective. The main contribution of this paper is development of a global linear solution which is based on the well-known Kruppa equations. We introduce a formulation different from the Huang and Faugeras constraints. The method has been extensively tested on synthetic and real data and promising results are reported.