Selection of Visual Descriptors for the Purpose of Multi-camera Object Re-identification - Feature Selection for Data and Pattern Recognition - page 275

D ,set u i and DEM M as the mean value over all subsets of u i normalized according
to the “best” value of M for all visual descriptors:

MD
i
max D (μ

MD
i

μ min D (μ )

RVI i ( D , M ) =

) ,

or RVI i ( D , M ) = . (12.28)

MD
i

MD
i

μ

RVI i (

, ) ∈[ , ]

D M 0 1

The formula for RVI is chosen based on the characteristics of DEM; if its value is
directly proportional to the expected results, i.e. the greater value of DEM the better
the descriptor is suited for visual object identification, then the first formula is used;
otherwise the second equation is employed.
Partial value-based ranking is aggregated in other to obtain the final value-based
ranking RV for the descriptor D :

M = 1

N M

i = 1

N U

RVP i (

, )

D M

RV ( D ) = , RV ( D ) ∈[ 0 , 1 ] . (12.29)

N M · N U

Position-based ranking, in contrast to the value-based one, takes into account only
integer position (place) of the DEM of the given descriptor against all other features.
For the given feature D ,set u i and DEM M , individual position-based ranking RPI
is defined as an integer value in the range

( N D —number of descriptor used)
that positions the descriptor as the best (1) or the worst ( N D ) among all descriptors.
Therefore, for a particular set and DEM, RVI values in the descending order and RPI
values in the ascending order create the same sequence of descriptors.
Partial position-based ranking is aggregated in other to obtain final, floating-point,
position-based ranking RP for the descriptor D :

[ 1 , N D ]

N U

M = 1

N M

i = 1

RPP i ( D , M )

RP ( D ) = , RP ( D ) ∈[ 1 , N D ] . (12.30)

N M · N U

12.6 Object Identification

The purpose of feature extraction presented in the previous sections is to employ
the gathered data on moving objects for the task of object re-identification. A typi-
cal scenario of the multi-camera system is that the object (e.g. a vehicle) leaves the
camera view and appears in another camera's view after some time. It is generally
not known in which camera the object will re-appear. Therefore, the problem may
be defined as follows. A set of feature vectors v i , j of the moving object O i , gathered
from each video frame j in which a given object was present, comprises a class .
For a newly appeared object in another camera, a new feature vector v is computed.