Accurate image alignment is needed for computing three-dimensional reconstructions from transmission electron microscope tilt series. So far, the best alignment results have been obtained by using colloidal gold beads as fiducial markers. If their use has not been possible for some reason, the only option has been the automatic 2D-cross-correlation-based registration methods. However, since the actual motion is three dimensional this approach is inaccurate and inappropriate for the whole problem. Conversely, we propose a novel method that uses the actual 3D motion model without any fiducial markers in the images. The method is based on matching and tracking corner features by fir st solving the underlying geometrical constraint of consecutive images in the tilt series. The results show that our method is competitive with the gold marker alignment in the level of accuracy and hence opens the way for new opportunities in the analysis of electron tomography reconstructions.