Precise Eye Detection Using Discriminating HOG Features

We present in this paper a precise eye detection method using Discriminating Histograms of Oriented Gradients (DHOG) features. The DHOG feature extraction starts with a Principal Component Analysis (PCA) followed by a whitening transformation on the standard HOG feature space. A discriminant analysis is then performed on the reduced feature space. A set of basis vectors, based on the novel definition of the within-class and between-class scatter vectors and a new criterion vector, is defined through this analysis. The DHOG features are derived in the subspace spanned by these basis vectors. Experiments on Face Recognition Grand Challenge (FRGC) show that (i) DHOG features enhance the discriminating power of HOG features and (ii) our eye detection method outperforms existing methods.