Dual Generative Models for Human Motion Estimation from an Uncalibrated Monocular Camera
讲座名称：
Dual Generative Models for Human Motion Estimation from an Uncalibrated Monocular Camera
讲座时间：
2008-12-25
讲座人：
樊国良
形式：
校区：
兴庆校区
实践学分：
讲座内容：
Dual Generative Models for Human Motion Estimation from an Uncalibrated Monocular Camera
Guoliang Fan, Associate Professor
School of Electrical and Computer Engineering
Oklahoma State University
Video-based human motion analysis is receiving increasing attention from computer vision and machine learning researchers. This research is motivated by a wide spectrum of applications and also fueled by the recent technical advances in several related areas. In this talk, we will discuss a generative model approach for human motion analysis where our goal is to estimate the 3-D gait kinematics from an image sequence captured by an uncalibrated monocular camera. There are two hypotheses in our research. One is that we can learn a low-dimensional non-linear space called gait manifold from a set of training gaits that provides a general gait representation either by its appearances or kinematics. The other is that an unknown gait can be synthesized in this space via non-linear interpolation of the training gaits. To address these two issues, we propose two tensor-based generative models for gait representation in both kinematic and visual spaces. Additionally, the concept of gait manifold is developed to represent different human gaits in a low-dimensional space. Particularly, we use the shortest path finding and spline fitting techniques in the tensor space to explore a 1-D gait manifold in each of the two generative models. A non-linear mapping is used to integrate the two gait manifolds, so that we can infer gait kinematics from gait appearances via two generative models. Moreover, a new particle filtering algorithm is developed for gait tracking and estimation where a segmental jump-diffusion Markov Chain Monte Carlo (MCMC) technique is embedded to accommodate the dynamic nature of gait variability in a long sequence. The proposed algorithm is trained from the CUM Mocap data set and tested on the Brown HumanEva data set, and the experiments show promising results compared to the state-of-the-art algorithms.
相关视频