Eye detection in unrestrained settings using efficient match kernels and SVM classification

Eye detection is a front end problem to be solved efficiently by face detection and human surveillance systems. Features such as accuracy and computational cost are to be considered for a successful approach. We describe an integrated approach that takes the ROIs outputted by a Viola and Jones detector, constructs HOGs features on those and learn an special function to mapping these features to a higher dimension space where the detection achieves a better accuracy. This mapping follows the efficient kernels match approach which was shown possible but had not been done for this problem before. Linear SVM is then used as classifiers for eye detection from those mapped features. Extensive experiments are shown with different databases and the proposed method achieves higher accuracy with low added computational cost than Viola and Jones detector. The approach can also be extended to deal with other appearance models.