Accurate human pose tracking using efficient manifold searching

In this thesis we propose novel methods for accurate markerless 3D pose tracking. Training data are used to represent specific activities, using dimensionality reduction methods. The proposed methods attempt to keep the computational cost low, without sacrificing the accuracy of the final result. Also, we deal with the problem of stylistic variation between the motions seen in the training and the testing dataset. Solutions to address both single and multiple action scenarios are presented. Specifically, appropriate temporal non-linear dimensionality reduction methods are applied to learn compact manifolds that are suitable for fast exploration. Such manifolds are efficiently searched by a deterministic gradient-based method. In order to deal with stylistic differences of human actions, we represent human poses using multiple levels. Searching through multiple levels reduces the effect of being trapped in a local optimal and therefore leads to higher accuracy. An observation function controls the process to minimise the computational cost of the method. Finally, we propose a multi-activity pose tracking methods, which combines action recognition with single-action pose tracking. To achieve reliable online action recognition, the system is equipped with short memory. All methods are tested in publicly available datasets. Results demonstrate their high accuracy and relative low computational cost, in comparison to state-of-the-art methods.