Real-Time Face Recognition from Surveillance Video - Intelligent Video Event Analysis and Understanding - page 162

The feature vector is extracted from the spatial relationships between local features.
It is therefore of very small dimensionality, allowing fast transmission over the net-
work and rapid matching to the database.
Our system was developed and tested using the following platforms and test data:

•

Face detection was implemented in C using the O PEN CV [5] libraries.

• Feature extraction algorithms were developed using M ATLAB R . The test data in-

cluded a live webcam stream, digital photos taken by the authors and the FERET
database of face images.

• Feature extraction was then ported to C++/O PEN CV. Test data included a live

webcam stream and videos of pose variation from CSIT 3 .

• The classification and matching system was implemented in C++ using the

O PEN CV SVM class.

Section 2 describes the design and architecture of our experimental system. Sec-
tions 3 and 4 describe the algorithms and operation of the front-end and back-end,
respectively. Our experimental results and conclusions are discussed in sections 5
and 6.

2 System Architecture

Our experimental system has been designed with the requirements of a real-time,
distributed sensor network in mind. It is envisaged that sensor nodes would be lo-
cated all around the Secure Corridor. Each node would contain a video camera, an
on-board processor and a network connection. The nodes would detect faces, then
send data to a back-end for matching to the database. The system architecture is
shown in figure 2.
First, faces are detected by the distributed sensor nodes. Feature extraction is also
carried out by the node. The high dimensionality of each video frame is reduced to
a feature vector of very low dimensionality. This vastly reduces the network band-
width required and spreads the processing load across the distributed network. Ob-
ject tracking could be carried out by the nodes, so that multiple face samples are
attributed to the same person. Feature vectors can then be tagged with an identifier
and sent to the back-end. (Object tracking was not implemented in our experimental
system).
The back-end system receives feature vector packets and unpacks them for pat-
tern matching with the database. This is treated as a classification problem, with
each individual enrolled on the system representing a class. Matching is performed
using a Support Vector Machine (SVM). Temporal relationships between frames
could be exploited to increase the likelihood of a correct match. In a production sys-
ten, the back-end would log the time and location (camera no.) of individuals under
surveillance, and would be configured to send alarms to operators.

3

Centre for Secure Information Technologies, Queen's University, Belfast

( http://www.csit.qub.ac.uk/ ). Thanks to Dr. Darryl Stewart and Adrian

Pass for providing the video database [24].