Selection of Visual Descriptors for the Purpose of Multi-camera Object Re-identification - Feature Selection for Data and Pattern Recognition - page 255

Chapter 12
Selection of Visual Descriptors
for the Purpose of Multi-camera
Object Re-identification

Piotr Dalka, Damian Ellwart, Grzegorz Szwoch, Karol Lisowski,
Piotr Szczuko and Andrzej Czyzewski

Abstract A comparative analysis of various visual descriptors is presented in this
chapter. The descriptors utilize many aspects of image data: colour, texture, gradi-
ent, and statistical moments. The descriptor list is supplemented with local features
calculated in close vicinity of key points found automatically in the image. The
goal of the analysis is to find descriptors that are best suited for particular task, i.e.
re-identification of objects in a multi-camera environment. The analysis is performed
using two datasets containing images of humans and vehicles recorded with different
cameras. For the purpose of descriptor evaluation, scatter and clustering measures
are supplemented with a new measure that is derived from calculating direct dissim-
ilarities between pairs of images. In order to draw conclusions from multi-dataset
analysis, four aggregation measures are introduced. They are meant to find descrip-
tors that provide the best identification effectiveness, based on the relative ranking,
and simultaneously are characterized with large stability (invariance to the selection
of objects in the dataset). Proposed descriptors are evaluated practically with object
re-identification experiments involving four classifiers to detect the same object after
its transition between cameras' fields of view. The achieved results are discussed in
detail and illustrated with figures.

· · ·

Keywords Video surveillance Image descriptors Multi-camera tracking Object

identification

12.1 Introduction

The popularity of video surveillance systems increases continuously along with
greater availability of solutions designed to ensure the safety of the monitored areas.
However, due to rapid development of multi-camera surveillance systems [ 17 ], they