Analysing the behaviour of small laboratory animals such as rats and mice is paramount in clinical research. We aim at recovering reliable anatomical measures of the skeleton of mice from videos and at demonstrating the robustness of these measures with quantitative validation. The most challenging aspects of this thesis reside in the study subject, mice, that is highly deformable, very fast and in the experimental conditions that do not allow for video data equivalent to what can be obtained with human subjects. In regards to the first challenge, we first focus on a marker-based tracking method with markers glued on the skin of the animal in Chapter 2. We show that the effects of the non-rigid mapping between skin and bones can be pre-empted by a weighting of the influences of the different markers in inverse kinematics. This allows us to demonstrate that, despite the non-rigid mapping, features on the skin of the animal can be used to accurately and robustly track the skeletal structures. We therefore develop a pipeline to process morphological data that leads to a generic animation model for the skeleton of mice. The weighted inverse kinematics method is validated with X-ray videos. Chapter 2 proves that points on the surface of the animal (on the envelope) can be used to track the skeletal structures. As a result, in Chapter 3, we propose a new deformation model of the envelope. This model, called OQR (Oriented Quads Rigging), is a flexible geometrical structure that has the nice deformation properties of cage-based animation. Like animation skeletons, OQR gives a high-level representation of the morphology and of the motion. We also show how, thanks to a well-deformed envelope, we can use a sub-set of the vertices of the deformed mesh as markers to apply the method of tracking of skeletal structures developed in Chapter 2. With Chapters 2 and 3, we have built a model of mice that allows us to animate at the same time the envelope and the skeleton. This model is parameterised by OQR. In Chapter 4, we therefore propose a method to recover the OQR parameters from either a sequence of meshes without temporal coherence or directly from segmented images. To regularise the tracking problem, we learn a manifold of plausible OQR configurations.
