Information Technology Reference
In-Depth Information
where C is a Boolean variable representing if there exists a correspondence be-
tween the tracks whose track link weight is w tr and track link count is m .If
insucient evidence has been collected then the track link weight and track link
count may not satisfy these conditions and the correspondence will not be found.
This will be the case when the camera network is first initialised and may also
happen for very short tracks.
Additionally, correspondences between tracks are declared following feedback
from foreign cameras. When a camera declares a correspondence between a pair
of tracks using equation 12, this correspondence is broadcast to all other cam-
eras. If this information implies that a track in the local camera and a track
in a foreign camera are the same real-world object, then a correspondence is
declared between these two tracks. This prevents asymmetry in the correspon-
dences across cameras, which would otherwise result from the fact that each
camera maintains its own set of tracking statistics.
Once a correspondence is found between two tracks, for objects from these
tracks that appear in future frames, the condition d s
d s max in equation 4 need
not hold for a linked pair of kernels to be initialised or strengthened between
their kernels. This ensures that statistics relating to the kernels in these two
tracks continue to be collected, for use in further multi-camera tracking.
3 Evaluation
3.1 Evaluation Datasets
Seven multi-camera video sequences were used as test sets for evaluation. The
evaluation test sets contained two or four cameras and comprised common FOV
scenarios, disjoint FOV scenarios, and hybrid scenarios containing some over-
lapping and some disjoint camera views. Test sets were recorded in an indoor
oce environment with uncontrolled lighting. Videos ranged in length from 2
minutes 50 seconds to 10 minutes 30 seconds, at approximately 10 frames per
second (fps) and at a resolution of 768
×
576 pixels or of 640
×
480 pixels.
3.2 Experimental Setup
Kalman Filter based single camera object tracking results were sanitised manu-
ally and used as input to the multi-camera system. Short tracks were automati-
cally removed as noise, even if they were not, for consistency.
The signature used in our evaluations divides the bounding box of an object
intoa4
4 grid and calculates luminance and hue histograms for the pixels in
each cell in this grid. Each histogram contains 8 bins, giving a total of 256 bins.
Euclidean distance was used to calculate the signature distance in section 2.1. We
selected this signature type as it is used by our single-camera tracking algorithm
in our overall object detection and tracking framework. However, it should be
noted that the multi-camera tracking algorithm presented in this paper could
be used with any signature type and distance metric.
×
 
Search WWH ::




Custom Search