In the wake of an increasing number of terrorist attacks, counter-terrorism measures are now a main focus
of many research programmes. An important issue for the police is the ability to track individuals and groups
reliably through underground stations, and in the case of post-event analysis, to be able to ascertain whether
specific individuals have been at the station previously.
While there exist many motion detection and tracking algorithms, the reliable deployment of them in a large
network is still ongoing research. Specifically, to track individuals through multiple views, on multiple levels
and between levels, consistent detection and labelling of individuals is crucial. In view of these issues, we have
developed a change detection algorithm to work reliably in the presence of periodic movements, e.g. escalators
and scrolling advertisements, as well as a content-based retrieval technique for identification.
The change detection technique automatically extracts periodically varying elements in the scene using Fourier
analysis, and constructs a Markov model for the process. Training is performed online, and no manual intervention
is required, making this system suitable for deployment in large networks. Experiments on real data shows
significant improvement over existing techniques.
The content-based retrieval technique uses MPEG-7 descriptors to identify individuals. Given the environment
under which the system operates, i.e. at relatively low resolution, this approach is suitable for short
timescales. For longer timescales, other forms of identification such as gait, or if the resolution allows, face
recognition, will be required.