A Rao-Blackwellized Mixed State Particle Filter for Head Pose Tracking in Meetings

This paper addresses the problem of head pose estimation in the context of meetings. More precisely, given a video of people involved in a meeting, the goal is to estimate the pose of people’s head with respect to the camera, which could ultimately translate into the estimation of the focusof-attention of people (who is looking at whom or what). To this end, we present a Rao-Blackwellized mixed state particle filter to achieve joint head tracking and pose estimation. Rao-Blackwellizing a particle filter involves splitting the state variables into two sets by marginalizing with respect to some of them, allowing for the exact computation of their posterior probability density function given the samples of the remaining state variables. This splitting and marginalization processes reduce the dimension of the configuration space to be sampled and lead to a more efficient particle filter requiring a lower number of particles to achieve similar tracking performance. To demonstrate this, we conducted experiments on a publicly available database consisting of people engaged in meeting discussions and for which the groundtruth is available thanks to the use of magnetic flock-of-birds sensors. The results from these experiments demonstrated the benefits of the Rao-Blackwellized particle filter model with fewer particles over the plain mixed state particle filter.