Head Tracking and Action Recognition in a Smart Meeting Room

Computer vision-based monitoring was used for automated recognition of the activities of participants in a meeting. A head tracker, originally developed for monitoring in a home environment, was evaluated for this smart meeting application using the PETS-ICVS 2003 video data sets. The shape of each person’s head was modelled as approximately elliptical whilst internal appearance was modelled using colour histograms. Gradient and colour cues were combined to provide measurements for tracking using particle filters. A particle filter based on Iterated Likelihood Weighting (ILW) was used which, in conjunction with the broad likelihood responses obtained, achieved accurate tracking even when the motion model was poor. It was compared to the widely used Sampling Importance Resampling (SIR) algorithm. Results are reported for tracking and recognition of the actions of the six meeting participants in the PETS-ICVS data. ILW outperformed standard SIR and reliably tracked all participants throughout the meeting scenarios.