Distance metric learning is a vital issue in people re-identification. Although numerous algorithms have been proposed, it is still challenging especially when the labeled information is few. Manifold regularization can take advantage of labeled and unlabeled information and achieve promising performance in a unified metric learning framework. In this paper, we propose Hessian regularized distance metric learning for people re-identification. Particularly, the second-order Hessian energy prefers functions whose values vary linearly with respect to geodesic distance. Hence Hessian regularization allows us to preserve the geometry of the intrinsic data probability distribution better and then promotes the performance when there is few labeled information. We conduct extensive experiments on the popular VIPeR dataset, CUHK Campus dataset and CUHK03 dataset. The encouraging results suggest that manifold regularization can boost distance metric learning and the proposed Hessian regularized distance metric learning algorithm outperforms the traditional manifold regularized distance metric learning algorithms including graph Laplacian regularization algorithm.