Human Motion Analysis: From Gait Modeling to Shape Representation and Pose Estimation
Abstract
This dissertation presents a series of fundamental approaches to the human motion analysis from three perspectives, i.e., manifold learning-based gait motion modeling, articulated shape representation and efficient pose estimation. Firstly, a new joint gait-pose manifold (JGPM) learning algorithm is proposed to jointly optimize the gait and pose variables simultaneously. To enhance the representability and flexibility for complex motion modeling, we also propose a multi-layer JGPM that is capable of dealing with a variety of walking styles and various strides. We resort to a topologically-constrained Gaussian process latent variable model (GPLVM) to learn the multi-layer JGPM where two new techniques are introduced to facilitate model learning. First is training data diversification that creates a set of simulated motion data with different strides under limited data. Second is the topology-aware local learning that is to speed up model learning by taking advantage of the local topological structure. We demonstrate the effectiveness of our approach by synthesizing the high-quality motions from the multi-layer model. The experimental results show that the multi-layer JGPM outperforms several existing GPLVM-based models in terms of the overall performance of motion modeling.On the other hand, to achieve efficient human pose estimation from a single depth sensor, we develop a generalized Gaussian kernel correlation (GKC)-based framework which supports not only body shape modeling, but also articulated pose tracking. We first generalize GKC from the univariate Gaussian to the multivariate one and derive a unified GKC function that provides a continuous and differentiable similarity measure between a template and an observation, both of which are represented by a collection of univariate and/or multivariate Gaussian kernels. Then, to facilitate the data matching and accommodate articulated body deformation, we embed a quaternion-based articulated skeleton into a collection of multivariate Gaussians-based template model and develop an articulated GKC (AGKC) which supports subject-specific shape modeling and articulated pose tracking for both the full-body and hand. Our tracking algorithm is simple yet effective and computationally efficient. We evaluate our algorithm on two benchmark depth datasets. The experimental results are promising and competitive when compared with state-of-the-art algorithms.
Collections
- OSU Dissertations [11222]