Multi-Shot Human Re-Identification Using Adaptive Fisher Discriminant Analysis

Human re-identification (re-id) in surveillance videos has been widely studied over the past few years. The research effort was mainly focused on building appearance models [1, 3] and learning a suitable metric to determine whether or not an image pair belongs to the same person[5, 8]. In terms of the available number of images per person in each camera view, the re-id problem can be categorized into single[7] and multi-shot [4] scenarios. In this paper, we propose a novel algorithm based on the Fisher criterion to learn a representative and discriminative feature subspace from image sequences to perform re-id. From the practical perspective of video surveillance applications, each person should have an image sequence available from tracking algorithms. To re-identify a person, we want to learn a feature space where images belonging to the same person stay close while images belonging to different people are far apart, which can be achieved by Fisher Discriminant Analysis (FDA) [2], defined as