Person Re-identification with Hyperspectral Multi-Camera Systems --- A Pilot Study

TL;DR

This study explores the potential of hyperspectral imagery to improve person re-identification in multi-camera systems by leveraging unique spectral signatures from skin, showing promising results over traditional RGB methods.

Contribution

It introduces a pilot hyperspectral re-identification approach, demonstrating its advantages over RGB images in challenging surveillance conditions.

Findings

Hyperspectral data significantly improves re-identification accuracy.

Spectral signatures from skin pixels are effective descriptors.

The approach outperforms traditional color-based methods.

Abstract

Person re-identification in a multi-camera environment is an important part of modern surveillance systems. Person re-identification from color images has been the focus of much active research, due to the numerous challenges posed with such analysis tasks, such as variations in illumination, pose and viewpoints. In this paper, we suggest that hyperspectral imagery has the potential to provide unique information that is expected to be beneficial for the re-identification task. Specifically, we assert that by accurately characterizing the unique spectral signature for each person's skin, hyperspectral imagery can provide very useful descriptors (e.g. spectral signatures from skin pixels) for re-identification. Towards this end, we acquired proof-of-concept hyperspectral re-identification data under challenging (practical) conditions from 15 people. Our results indicate that hyperspectral data result in a substantially enhanced re-identification performance compared to color (RGB) images, when using spectral signatures over skin as the feature descriptor.