Information Technology Reference
In-Depth Information
Fig. 3.
i LIDS multiple camera tracking scenario definition provided by HOSDB
identifier (
k
∗
) of the object in the wide area is based on its previous occurrence
(spatio-temporal,
o
) and its state (appearance,
s
):
k ∗
(
o, s
)=
argmax
k
P
(
k|o, s
)
≈ P
(
o|k
)
P
(
s|k
)
(1)
Because of this, we must (approximately) know the camera topology. The figure
3 is suitable enough for the learning step. We have used annotations provided
by the HOSDB on i-LIDs MCT dataset. There are 5 cameras and several areas
from where a new object can enter.
The object appearance and bias is automatically learned (or summarized)
using Gaussian mixture models [8] or optionally SVM. The probability
P
(
s|k
)
is then determined by a similarity search (the distance is normalized using the
sigmoid) with respect to the expected bias, which is simply subtracted.
4.4 Querying
The SUNAR queries are of two types - on-line used for instantaneous condition
change and especially for identity preservation as described above; and off-line
queries, able to retrieve all the metadata from processed camera records in the
wide area after an accident, crime or a disaster happens.
We can distinguish two types of operations: environmental and trajectory op-
erations. Environmental operations are relationships of an object's trajectory
and a specified spatial or spatio-temporal environment, such as enter, leave,
cross, stay and bypass [2,5]. Trajectory operations look for relationships of two
or more trajectories restricted by given spatio-temporal constraints, such as to-
gether, merge, split and visit.
We have also implemented
2
similarity queries based on MPEG-7 features in
the PostgreSQL database as a vector (array) distance functions - Eukleidean
(Mahalanobis), Chebyshev and Cosine distance.
2
Avaiable at
www.fit.vutbr.cz/research/view_product.php.en?id=73
Search WWH ::
Custom Search