Person Re-identification with Soft Biometrics Through Deep Learning

Abstract

Re-identification of persons is usually based on primary biometric features such as their faces, fingerprints, iris or gait. However, in most existing video surveillance systems, it is difficult to obtain these features due to the low resolution of surveillance footages and unconstrained real-world environments. As a result, most of the existing person re-identification techniques only focus on overall visual appearance. Recently, the use of soft biometrics has been proposed to improve the performance of person re-identification. Soft biometrics such as height, gender, age are physical or behavioural features, which can be described by humans. These features can be obtained from low-resolution videos at a distance ideal for person re-identification application. In addition, soft biometrics are traits for describing an individual with human-understandable labels. It allows human verbal descriptions to be used in the person re-identification or person retrieval systems. In some deep learning based person re-identification methods, soft biometrics attributes are integrated into the network to boot the robustness of the feature representation. Biometrics can also be utilised as a domain adaptation bridge for addressing the cross-dataset person re-identification problem. This chapter will review the state-of-the-art deep learning methods involving soft biometrics from three perspectives: supervised, semi-supervised and unsupervised approaches. In the end, we discuss the existing issues that are not addressed by current works.