Learning to describe human activity and body pose from images
Date: Wed 23rd June
Location: CS 338
Speaker: Dr Stephen McKenna, University of Dundee
The first part of this talk presents a vision system for a supportive home environment. Its aims are automatic
semantic summarisation of human activity and detection of unusual inactivity. A method is presented for automatically
learning a context-specific spatial model in terms of semantic regions (inactivity and entry zones). Penalised
likelihood functions, incorporating prior knowledge of the size and shape of semantic regions, are used to estimate
Gaussian mixtures with MDL. This encourages a one-to-one correspondence between mixture components and regions. The
resulting model enables human-readable summaries of activity to be produced and unusual inactivity to be detected.
Results are presented using overhead sequences tracked with a modified particle filter.
The second part of the talk addresses the task of estimating human body pose given a single image. A model of human
appearance is presented that copes well with wide variations in illumination, clothing, occlusion by other objects,
self-occlusion and background scenes. The likelihood model is based on learned feature divergence distributions and
combines probabilistic body part models, inter-part constraints and paired-part constraints. Partial configurations
with differing numbers of instantiated body parts are compared based upon learnt likelihood ratios. Some results from
applying an optimisation scheme with this model will be presented.