Video-based human motion estimation has recently received great interest due to its wide applications and has been advanced by the recent progress in the fields of computer vision and machine learning. We are interested the estimation of human body configurations from image sequences taken by an uncalibrated monocular camera. Specifically, we focus on gait estimation that is very useful for biometrics and biomechanical modeling applications. One key issue is the dimensionality reduction that would reduce the data redundancy and explore the intrinsic non-linear low-dimensional structures among various gait kinematics and appearances. We propose generative model-based approaches that provide a unified and general gait representation in both kinematic and visual spaces. The new method achieves the state-of-the-art results on estimating the gait kinematics.

(Left: the observed gait silhouettes, the synthesized gait silhouettes, the comparison between real/estimated gait kinematics. Right: the two examples of training sequences)

Related Publications

• X. Zhang and G. Fan "Dual Generative Models for Human Motion Estimation from an Uncalibrated Monocular Camera", in Proc. International Conference on Pattern Recognition (ICPR), Tampa, Florida, Dec. 2008.