Region Constraint Person Re-Identification via Partial Least Square on Riemannian Manifold

Person re-identification refers to matching images of pedestrians from camera networks distributed at different locations. It involves multiple challenging problems, especially large variations in illumination, viewpoint, and occlusion which exert negative affect on recognition rate. In this paper, we propose the partial least square (PLS) algorithm based on the Riemannian manifold. The covariance matrix is established by extracting the feature information of the region of the torso and leg images of a pedestrian. To find a suitable tangent plane in the covariance matrix of the Riemannian manifold, we create a relation between the torso and leg image by using the PLSs algorithm to obtain the correlative feature information of pedestrian images. Using unsupervised learning, this method performs well in the single-shot scenario. Experimental results on publicly available benchmarks show better performance of the proposed method.