Object Tracking over Multiple Uncalibrated Cameras Using Visual, Spatial and Temporal Similarities - Advanced Concepts for Intelligent Vision Systems - page 171

In addition, the transit time, t ,givenby,

t = f local − f foreign (7)

is computed, where f local and f foreign are the frames in the local and foreign track
snapshots respectively. This provides a measure of the time it took an object to
move from the real-world location in the foreign camera to the real-world location
in the local camera.
Each linked pair of kernels is associated with the most recent signature weight,
w s , the kernel link weight, w k , and a history of transit times. The kernel link
weight is used later in location-based kernel matching and when correspondences
between tracks are found in section 2.3. The historical set of transit times are
used later in computing temporal similarities in section 2.2. In addition, each
linked pair of kernels is associated with a history of the local and foreign track
snapshots each time the link was strengthened.

Linking kernels using the location-based kernel matching method.
Chan-ges in camera angle often affect the apparent pose of an object as viewed by
the camera. As a result, different parts of the object may have different colours
visible to different cameras. This leads to large signature distances when using
the signature-based kernel matching method just described, which prevents ker-
nels of tracks corresponding to the same real-world object from being correctly
linked. The location-based kernel matching method described in this section
addresses this problem. A history of linked pairs of kernels must have already
been established before location-based kernel matching can be performed. This
can be done by running the STAC algorithm using only signature-based kernel
matching for some time before enabling location-based kernel matching.
The proposed location-based kernel matching is performed after receiving for-
eign track snapshots from other cameras, in parallel to the signature-based kernel
matching method. For each received foreign track snapshot, of the set of histor-
ical linked pairs of kernels containing the kernel in the foreign track snapshot
and a kernel in the local camera, the pair with the greatest kernel link weight
is identified. A historical linked pair of kernels is any linked pair of kernels ini-
tialised in a previous frame. If the local track passed through the local kernel in
this pair in a previous frame, then the current signature of the locally tracked
object is replaced with its signature from this previous frame. Following this, the
signature distance between the new local signature and the signature in the for-
eign track snapshot is calculated, and if it is below the threshold d s max , then the
signature weight, kernel link weight and transit time are initialised or updated
as per equations 4-7.

Selecting the best linked pairs of kernels. The kernel matching process
described until this point will result in a set of linked pairs of kernels between
local kernels that the locally tracked object passed through and kernels in for-
eign cameras. For a given locally tracked object, if there exists more than one
linked pair of kernels between the local camera and a foreign camera, only the