Re-identificación de personas mediante la distancia de Mahalanobis

Résumé

Person re-identification requires the learning of a distance metric able to compare two images and decide if they belong, or not, to the same person. The automation of this task, in order to be applied in intelligent video-surveillance, involves a great challenge, due to the presence of people with similar appearance. For that reason, it is necessary to learn discriminative features and a metric to properly combine them. However, the variations of illumination, perspective, background, resolution and scale between two images of the same person, which were captured from different views, make his or her apperance vary, hampering the re-identification. This paper proposes coding the view-to-view tranformations in a Mahalanobis matrix, whose estimation has been integrated into the discriminative features learning. In that way, these features can render the dissimilarity mainly due to appearance changes intead of the view changes. This estimation has been implemented as a new layer of a deep convolutional neural network, which has been trained and tested over the PRID2011 dataset