We propose a method for the non-rigid tracking of small animals in video sequences. Based on an image sequence showing the animal, first a sub-set of images with coherent pose is chosen automatically. Then a robust rigid registration determines the coarse animal body pose in the set of frames and a subsequent non-rigid group-wise registration using the minimum description length principle learns a deformation model of the animal. It obtains a non-rigid mapping between the individual positions and shapes of the animal during the sequence, based on discrete sets of landmark candidates and according texture features. This is of high relevance in small animal research to integrate signals from multiple frames, and cannot be achieved by standard continuous registration methods. We report first experimental results on video sequences of a rat.