Online appearance learning for 3D articulated human tracking

A human appearance modelling framework where colour distributions are associated with surface regions on an articulated body model is presented. In general, these distributions are unknown, multi-modal and changing in time. We therefore propose using recursively updated histograms to represent them. For a certain pose, a set of histograms may be collected and a likelihood constructed based on the histograms' similarity with the previously learned histograms. To ease histogram estimation and improve computational efficiency, a merging and splitting algorithm is derived which groups surface regions based upon histogram similarity and prior knowledge of clothing layout. An investigation of the behaviour of this likelihood shows it to be broad, smooth and peaked around the correct location, a good candidate for coarse sampling and gradient-based search methods. We show how conditioning the likelihood to maximise foreground usage reduces secondary maxima. Finally, we present results from tracking a challenging sequence.