We use Psyclone to design and build a Computer Vision architecture for Cognitive Tracking of Object. We use the SDK to create both low and high level Vision Modules to process the incoming video which is streamed through the Whiteboard MediaStreams. The processing is prioritised so lower level modules are guaranteed to get enough time to do their work.

The video is initially analysed to find image features such as corners, edges and motion and the perceived information is gathered centrally in a large mathematical representation for further analysis. The higher level modules have concurrent access to all the data and can even modify parts of it while others analyse other parts.

Psyclone uses very fast streaming and caching of Media data between modules and even between computers when distributing the system across a network. We use the built-in PsyProbe interface to monitor the system while it is running and we use the Psyclone GUI for visualisation of multiple processes simultaneously.

We are able to seamlessly mix C++ and Java modules, running internally or externally to Psyclone. All of this is manages with one single config file which fully specifies the behaviour, data paths and processing of Modules, Whiteboards and the system as a whole.

Using simple modules which subscribe to messages and binary data it is easy to create the proper data flow and allow this to be automatically regulated to adapt to new scenarios. It is also easy to see the flow in PsyProbe and make sure that each module is doing the right thing.

It is our plan to implement some of the lower level Vision Modules to make use of many low-cost NVidia graphics cards for the heavy image processing and the Psyclone modules are ideal for this purpose as they completely wrap all communication with Psyclone so each can concentrate on its input, output and processing task.

The system very easy to regulate and using priorities we hope to be able to reach a stable online and real-time architecture for Tracking in Computer Vision.