Dense Sigma set representation for pedestrian detection

This paper proposes an effective and efficient pedestrian detection method which is based on 2nd order statistics image representation and linear Support Vector Machine classifiers (SVMs). After arranging a sample image into a dense structure of regions, the key issue is how to represent each image region. Exploiting the traditional covariance matrix region descriptor is a feasible way. However, it requires the usage of computationally demanding non linear classifiers, such as the radial basis function kernel SVMs, because the distance metric between covariance matrices is not in Euclidean space. Instead, we propose Sigma set as the region descriptor to represent each image region. As a result, this representation, which is named dense Sigma set, together with the efficient linear classifier achieves high accuracy. Experimental results on the INRIA pedestrian detection dataset show that dense Sigma set is a powerful and discriminative image representation method for pedestrian detection.