为了降低群体动画中生成大量自然而又相似的人体运动的难度和复杂性，研究了一种基于学习的群体动画生成技术。该技术首先通过建立基于高斯过程隐变量模型和隐空间动态模型的运动姿势学习模型，将高维运动姿势映射到低维隐空间中，并在低维隐空间对相邻姿势的动态演化进行建模；然后通过对已有运动数据的学习来获得组成该运动的姿势的概率分布，再通过隐空间中的动态预测和Hybrid Monte Carlo采样来得到符合给定概率分布的隐轨迹；最后通过姿势重构来得到与原运动非常相似但又不同的一系列自然的运动，以产生群体动画，从而避开了传统的基于几何和物理约束的逆运动方法固有的困难和复杂性。

A group animation generation method based on machine learning was proposed in order to reduce the complexity of generating mass of similar but different natural human motions in group animations. There are two models. Poses learning model was built based on Gaussian process latent variable model to characterize a specific motion and dynamic model was built in latent space to characterize the dynamic evolving process of neighboring poses in latent space. These models can be represented as probability distribution over all poses composing the motion by learning from existing motion data. Dynamic prediction can be made in latent space for giving initial state, then hundreds of latent trajectories by Hybrid Monte Carlo sampling according to given probability distribution can be obtained. Group animations can be implemented by generating a series of similar but different natural motions reconstructed from these latent trajectories, thereby avoid the difficulty and complexity of calculating geometric relationship and physical constrains in inverse kinematics.