Modélisation anatomique utilisateur-spécifique et animation temps-réel. Application à l'apprentissage de l'anatomie. (User-specific real-time registration and tracking applied to anatomy learning)

titre : User-specific real-time registration and tracking applied to anatomy learning. To make the complex task of anatomy learning easier, there exist many ways to represent and structure anatomical knowledge (drawings, books, cadaver dissections, 3D models, etc.). However, it may be difficult from these static media to understand and analyze anatomy motion, which is essential for medicine students. We introduce the "Living Book of Anatomy" (LBA), an original and innovative tool to learn anatomy. For a specific user, a 3D anatomical model (skin, skeleton, muscles and organs) is superimposed onto the user’s color map and animated following the user’s movements. We present a real-time mirror-like augmented reality (AR) system. A Kinect is used to capture body motions. Three challenges have been tackled : The user-based model deformation challenge identifies the user’s body measurements and use it to register our 3D anatomical model. We propose and evaluate two different registration methods. The first one is real-time and use affine transformations attached to rigid frames positioned on each joint given by the Kinect body tracking skeleton. This allows to deform the 3D anatomical model using skinning to fit the user’s measurements. The second method needs a few minutes to register the anatomy and is divided in 3 parts : skin deformation (using Kinect body tracking skeleton and the Kinect partial point cloud) ; combination with strict anatomical rules we register the skeleton ; soft tissue deformation to fully fill the space inbetween the registered skeleton and skin. Through the real-time model animation challenge, we want to capture realistically and in real-time the user’s motion. Reproducing anatomical structure motion is a complex task due to the noisy and often partial Kinect data. We propose here the use of anatomical rules concerning body joints (angular limits and degrees of freedom) to constrain Kinect captured motion in order to obtain plausible motions. A Kalman filter is used to smooth the obtained motion capture. The Augmented Reality model visualisation challenge, is about embedding visual style and interaction, using a full body reproduction to show general knowledge on human anatomy and its differents joints. We also use a lower-limb as structure of interest to higlight specific anatomical phenomenon, as muscular activity. All these approaches have been integrated in a working system detailed in this thesis. This has been validated through live demos during different conferences and through preliminary user studies.