Information Technology Reference
In-Depth Information
vehicle) detected by the method described in the previous section. The de-
scriptors that we use are correspond to the energie calculated on every sub-
band, by a decomposition in wavelet of the optical flow estimated between
every image of the sequence. We obtain a vector of 10 bins, they represent
for every image a measure of activity sensitive to the amplitude, the scale
and the orientation of the movements in the shot.
6 Experimental Results
Experiments are conducted on the many sequence from TRECVid2009 database
of video surveillance and many other sequences from road tracs. About 20 hours
are used to train the feature extraction system, that are segmented in the shots.
These shots were annotated with items in a list of 5 events.We use about 20
hours for the evaluation purpose. To evaluate the performance of our system we
use the common measure from the information retrival community: the Average
Precision. Figure 6 shows the evaluation of returned shots. The best results are
obtained for all events.
Fig. 6.
Our run score versus Classical System (Single SVM) by Event
7Conluon
In this paper, we have presented preliminary results and experiments for high-level
feature extraction for video surveillance indexing and retrieval. The results ob-
tained so far are interesting and promoters.The advantage of this approach is that
allows human operators to use context-based queries and the response to these
queries is much faster. The meta-data layer allows the extraction of the motion
and objects descriptors to XML files that then can be used by external applica-
tions. Finally, the system functionalities will be enhanced by a complementary
tools to improve the basic concepts and events taken care of by our system.
Acknowledgement
The authors would like to acknowledge the financial support of this work by
grants from General Direction of Scientific Research (DGRST), Tunisia, under
the ARUB program.
Search WWH ::
Custom Search