Contest description

Nowadays, it is common to use CCTV systems for monitoring infrastructures such as airports, railway stations and shopping centres. The environment is often monitored by multi-camera systems. Over the last two decades, an increasing interest has been shown in the field of automated visual surveillance. This has led to the design of a plethora of automated video surveillance systems. Considerable efforts have been spent to develop algorithms analysing image data coming from such systems. Today we notice that it is still a challenge to design Computer Vision algorithms for such analysis. One reason is the wide baseline setting in such camera systems where the non-overlapping fields of view are often a result of these wide baselines.

The contest consists of two challenges:

  • The first one consists of tracking people across the cameras having wide baselines and non-overlapping fields of view in a consistent and accurate way through the whole system.
  • The second challenge consists of discovering the camera network topology.

In a wide baseline camera network configuration, the cameras see the object from totally different viewpoint and totally different illumination conditions. One of the main difficulties consists of solving the problem of wide baseline camera setting considering also the case of non-overlapping fields of view between the cameras. Another problem to be solved is how the object to be tracked can be represented in order to be able to track objects consistently and accurately. Lastly, the recognition of the camera network topology is another problem to be faced. After initialising the objects to be tracked (no restriction on either manual or automatic initialisation), the whole process has to run automatically, which means no more user interaction or use of any new static information (for example, a new manual selection) are possible.

No severe restrictions on the approach are applied by the contest to solve the challenges; this means solutions using any approach (e.g. appearance-based, motion-based, biometric feature-based, etc.) are possible. Solutions based on learning on-line the visual representation of the object are encouraged.

For a pdf version of the contest call for participation, please click here