SUNAR Surveillance Network Augmented by Retrieval - Advanced Concepts for Intelligent Vision Systems - page 159

4.1 Metadata Cleaning

The preprocessed data is supposed to be incomplete or duplicate, biased and
noisy. Thus, moving objects are modeled as dynamic systems in which the
Kalman filter optimally minimizes the mean of error [5] and it can fill in the
missing information (position and velocity) for a few seconds in case the object
has been occluded, for instance 1 .
At the cleaning step, SUNAR stores metadata representing moving objects
and information about the environment under surveillance.

4.2 Indexing and Storing

The database model consists of three database schemes in the SUNAR database -
Process, Training and Evaluation according to their purpose. All schemes contain
three main tables that correspond to the fundamental concepts - Object, Track
and State (as in our former work [5]). Object is an abstract representation of a
real object (having a globally unique ID), it is represented by its states. A state
consists of two types of features - visual properties (as described in section 3) and
spatio-temporal features. The latter are represented by location and velocity of
an object at a moment. A track is a sequence of such states in a spatio-temporal
subspace of the area under surveillance followed by one camera.
The training scheme contains also tables containing statistics and classifica-
tion models according to the method used. For instance, a simplified Bayesian
model table contains columns for source and destination camera IDs, in which
objects are passing through. Next columns represent the number of training
samples, a prior probability, averages and variances of handover time, trajectory
states and visual features. Trajectories are summarized as a weighted average of
cleaned states, where the weight is highest at the end of the trajectory. If cameras
are overlapping, the handover time may be negative. The average and variance of
different feature descriptors acts as the visual bias removal (illlumination, color,
viewpoint and blob size calibration) for the integration step.

4.3 Multiple Camera Integration

The training schema described before is rather simplified. In fact, we use Gaus-
sian Mixture Model and Support Vector Machine [14,8] models of the (inverted)
Kalman filter state as described in our previous work [5]. The inverted state is
computed using Kalman filter in the opposite direction the object moved through
one camera subspace followed by one camera. The goal of this trick is the clas-
sification of the previous subspace (camera) in which it was seen last time most
probably.
The object identification then maximizes the (prior) probability of a previ-
ous location (camera) multiplied by the normalized similarity (feature distance
without bias) to previously identified objects according to average time con-
straints and visual features in the database [5,10]. More formally an optimal

1 Available at www.fit.vutbr.cz/research/view_product.php.en?id=53