Dadgar, S. Amin (2009) 3D human pose estimation from silhouettes in activity constrained scenarios. (MSc(R) thesis), Kingston University, .

Abstract

The goal of this thesis is to introduce the main framework to estimate the 3D pose of human body. The main contribution of this work is the Dense Gaussian Mixture Modelling of the training set initialised by information extracted from the gait cycle. The method proposed in this work is a generative approach based upon silhouette matching which learns a specific activity (e.g. walking, jogging) to bias the inference toward the most probable configurations. First of all Principal Component Analysis is used to reduce the dimensionality of the state space. Next, motions were segmented to elementary movements using the information extracted from gait cycle. This helps us to increase the number of Gaussians in the GMM model effectively and shows a great improvement compared to lower number of clusters extracted within other framework such as Expectation Maximisation. The temporal dependencies between these elementary movements are then captured by a Hidden Markov Model. Annealed Particle Filter is used to resolve the issue of local maxima which caused the solution stuck in the center of each Gaussian in the inference phase. The metric we used to return the best pose is the Chamfer distance and shows reasonable performance in this application of 3D pose estimation. Results for walking and jogging are presented and shows they are comparable with state-of-the-art. The approach can be considered as a basic framework of this application to be developed and enhanced by further research.

Actions (Repository Editors)

Item Control Page