Abstract

We address the problem of re-identifying a person in videos captured from cameras with disjoint field of view in different background. The varying pose, scale, and motion pattern of a person makes this task challenging. In this work we propose two frameworks wherein the first one exploits only the appearance information whereas the second incorporates motion information as well. The first framework utilizes Support Vector Machine (SVM) and inner product for measuring the similarity between query and gallery videos. In the next framework we present a neural network based system for re-identifying persons in videos. We employ an LSTM as a classifier and train it over output vectors from the memory cells corresponding to different persons obtained from another LSTM. Both the proposed methods outperform existing state-of-the-art methods.