Information Technology Reference
In-Depth Information
Fig. 4.
Illustration of the multiple camera tracking process of the SUNAR system
including manual ground truth annotations provided by HOSDB and NIST
4.5 Analysis
We perform several types of video analysis, mainly classification and clustering as
illustrated in figure 1b. The first type is the modeling based on visual appearance
of an object (color layout, blob) using Gaussian Mixture Models (GMM, [8]).
Second, we perform trajectory classification based on Gaussian Mixture Mod-
els as needed for the multiple camera identification as in section 4 and Hidden
Markov Models (HMM). In the article [11] we selected few scenes, where some
easily recognizable human behavior occurs. For example, one concept represents
if people go through turn pikes or not. The HMM are trained on such classes. The
trajectory which doesn't fit any HMM model (with respect to some threshold)
is considered to be abnormal. In addition, SUNAR uses velocity and accelera-
tion as training features, which describe and discover some abnormalities better
(jump over).
Moreover, using the spatio-temporal queries, we can discover splitting and
merging objects, opposing flow (together with GMM and aggregate functions)
or an object put (operations enter, split, leave and stay).
5 Evaluation
The previous evaluations such as Performance Evaluation of Tracking and Surveil-
lance (PETS [17]) dealt with other aspects of computer vision than multiple cam-
era surveillance with non-overlapping camera fields of view. They either dealt with
classical single camera tracking or they have concerned more on the event detec-
tion as Classification of Events, Activities, and Relations. For instance, events
so-called left baggage, split, hug, pointing, elevator no entry are detected in the
TRECVid Surveillance Event Detection evaluation [19].
The AVSS 2009 Multi-Camera Tracking Challenge [20] was the first evaluation
campaign that used the annotated Multiple-camera Tracking (MCT) Dataset
Search WWH ::
Custom Search