Learning large margin likelihoods for realtime head pose tracking

We consider the problem of head tracking and pose estimation in realtime from low resolution images. Tracking and pose recognition are treated as two coupled problems in a probabilistic framework: a template-based algorithm with multiple pose-specific reference models is used to determine jointly the position and the scale of the target and its head pose. Target representation is based on histograms of oriented gradients (HOG): descriptors which are at the same time robust under varying illumination, fast to compute and discriminative with respect to pose. To improve pose recognition accuracy, we define the likelihood as a parameterized function and we propose to learn it from training data with a new discriminative approach based on the large-margin paradigm. The performance of the learning algorithm and the tracking are evaluated on public images and video databases.