Once the camera is set up and connected to the user's console computer that is running Cobalt, the user has to describe to Cobalt what he/she wants to achieve. This is done simply by defining areas of interest in the scene and possibly imposing conditions about target size and and direction of motion. The user is not concerned with weather, shadows, or any such interference due to intrinsic changes in the scene.

In this way, user training is minimised: under most circumstances, simply drawing a box or two on hte screen will offer effective event detection. That is what we call "zero learning". There is, however, ample scope for the trained user to deal with specfic detection criteria.

The examples on this page illustrate this in realistic circumstances.

Cobalt currently "learns" about a scene within the space of a few frames, ie. within a seoond or two of starting up or switching to a new scene. There is no need for the user to tell Cobalt anything about the scene, except to "describe" what is to be detected. Thereafter Cobalt adapts in real time to any global change in the scene that might effect the detection of an event in a specified area.

Such global changes might be variations in light due to scudding clouds, camera shake, trees swaying in hte wind and casting moving shadows. At night, the global change might simply be due to "noise" in the image that is due to a lack of sufficient lighting.

Cobalt achieves this level of being able to filter out "background" and "foreground" interference through building a mathematical model of the scene and adjusting the model as the frames are acquired. With modern processors the Cobalt algorithms can do this at the full frame rate.

The Cobalt mathematical algorithms are closely related to AI techniques, though they differ in their detailed structure. When Cobalt is started up there is an automated "learning" or "training " phase lasting a second or two during which time the mathematical model sets up its basic parameters. No user help is required in this: it is in this respect an AI system that is pre-programmed to decide on cetain factors that effect image handling.

The video stream generated by Cobalt software consists of two components: the compressed (and perhaps encrypted) video stream and an Image-DNA stream. The mathematical models referred to above operate using the image DNA which encodes the detail about the scene. The parameters that define the model are computed at high speed so as to minimise the interference from extraneous influences at the full frame rate.

The major impact of neural net technology or, for that matter any other AI technology, is in the arena of object classification/recognition. The AI functionality depends very much on the application to which Cobalt is put. In scene surveillance, where it is necessary to make decisions in circumstances where an object is not fully perceived, we might be concerned with recognition on a rainy night, or in a sitaution where it is partially obscured. In other circumstances, as in process monitoring, a decision may have to be taken if marginal evidence for some critical event arises.