Accelerating statistical human motion synthesis using space partitioning data structures

Statistical model-based motion synthesis enables the constrained generation of natural human motions. In this paper, we present a combination of statistical model-based motion synthesis with a motion retrieval method. We fit a statistical motion model to the low-dimensional latent space obtained by applying principal component analysis to a functional representation of example motion data. In order to accelerate the synthesis, we construct a space partitioning tree on a dense set of samples from the model projected into a feature space. At run-time, the tree is traversed based on an objective function to find the best sample fitting user-defined constraints. After finding a good initial guess, we apply numerical optimization to reach constraints exactly. We evaluated the accuracy and efficiency of the search in the tree using different space partitioning methods on synthetic and motion capture data. In our experiments, the most stable and accurate search results were achieved using the k-Means++ algorithm for the tree construction.