Head pose estimation based on steerable filters and likelihood parametrized function

In this paper, we propose a new discrete head pose estimator that combines appearance template and discriminative learning. Our approach consists on constructing a reference model for each considered head orientation. To do this, we first locate the head patch using a skin color based filter. The reference models are then elaborated from steerable filters which are applied in order to extract feature vectors. We chose to apply such filters since they are robust to global geometric deformations and view point changes. Next, we learn parameters of likelihood function from training data with a discriminative approach. When a new image is considered, a feature vector based on steerable filters is extracted from the localized head patch. Subsequently, head pose is estimated using likelihood parametrized function. The performance of our estimator is evaluated on PRIMA-POINTING database showing that the proposed approach is very competitive compared to other existing methods.