Output details

Tracking human pose with multiple activity models

<23>Core work from a Ph.D. project supervised by Dr. Li (successfully defended April 2010), describing a system which tracks, in 3D, human poses from synchronised videos. System operates at reduced computational cost when tracking previously seen activities but can also “work harder” to recover new, previously unseen behaviours. The system provides efficiency savings in any non-invasive motion tracking scenario (e.g. surveillance). At a theoretical level the work contributes a method for fairly combining two or more Bayesian inference tasks with different levels of complexity. Thorough evaluation on large, publicly available datasets highlights the robustness of the approach.