Abstract: We present a method for camera autocalibration from images of an unknown planar object. The few existing methods for that problem proceed by non-linear batch optimization, whereas our method is of sequential nature, being based on Kalman filtering. Its overall computational cost is much smaller, thus enabling the use of video sequences as input for real-time, on-the-fly, autocalibration. Being able to use video efficiently is very beneficial, since a stable autocalibration requires images taken with significantly varying camera orientations : batch methods put limits on the number of images used, hence require either to select key frames from a video, or to use discrete image sequences, for which the matching is more difficult. Our method is not concerned by these issues. Although not optimal (in the bundle adjustment sense), our method compensates by using indeed all available information in an image sequence. Experimental results show that it performs similarly to a global batch method, in terms of accuracy and convergence, while allowing real-time autocalibration in practice.