With our research into activity recognition we enable smart cameras to realize activities of special interest. Examples of such activities are:

• Falling person
• Wrong way motion
• Crowd forming

The desire for designing the activity recognition as flexible and adaptive as possible requires the use of leading edge machine learning algorithms. No matter how easy programming becomes, our machine learning based approach leads to smart cameras capable of being adaptable to applications with no programming needs at all - just by presenting example situations.

One of the most interesting benefits of activity recognition in a very early state of the processing is the virtually unlimited expandability in terms of smart cameras and the gain of computing power on the server side. With this approach the connected server is able to concentrate on solving higher level problems.

The classification of objects is required in a various range of vision applications. Our SmartSurv system needs to distinguish between the tracked objects, e.g., cars vs. persons or adults vs. infants. Depending on the classification results, it is possible distinguish between the activities that should be recognitized.

A more traditional application for object classification based purely on appearance, e.g., the SmartFruit camera presented on the VISION2006 at Matrix Vision booth. Note that the classification runs fully embedded within the smart camera.

Our SmartSurv system is also capable of integrating the results of multiple of such classification cameras.

We are working on a toolbox of classification instruments based on recent machine learning algorithms customizable for specific applications with an easy to handle userinterface.