Study On Generating Reactive Motions For Human Animation

With the increasing requirement for human animation in virtual reality, 3D games and digital entertainment, the motions of a virtual human are not only required to be generated in a natural and controllable way, but also expected to respond to unexpected perturbation realistically. There are three key problems to create reactive human animation: how to improve the physical realism of human motion and make the character's behavior more life-like, how to exactly respond to the strength and direction of external impulse and how to generate real-time reaction. This dissertation focuses on these problems on balance maintenance for external perturbation during quiet standing, data-driven reactive motion synthesis, hybrid control based reactive motion generation and balance maintenance for external perturbation during locomotion.On balance maintenance for external perturbation during quiet standing, this thesis proposes to analyze and calculate the balance from both sagittal plane and frontal plane. According to research results of biomechanics, we construct a physical model of human body based on double inverted pendulum. Proportional-derivative controller is used to approximate internal muscle actuation and drive human to move. Under the hip and ankle control strategies, the center of mass of a virtual human is kept in the support area, so that the virtual human can counteract external force by swinging body, and then recover balance. A heuristic method is used to tune the gains. It speeds up the tuning process tremendously.On the synthesis of data-driven reactive motion, this thesis proposes a search algorithm to select a matching reactive motion to respond external perturbation by the combination of momentum perturbation and pose comparison. Traditional motion synthesis algorithms only select a matching motion from the database according to the distance between two poses and neglect the change of momentum. Therefore the result is not realistic and costs a lot of computation. The algorithm given in this dissertation utilizes the similarity of momentum changes between a real person and the synthetic virtual human after an impact. Therefore it can provide more realistic effects. Approximate nearest neighbor (ANN) search algorithm is used to speed up the search process and a skeleton LOD is used to reduce the dimensionality of motion data. Experiments show that our algorithm can simulate the reactive motion of a virtual human with quick response to the user's input virtual force very well.