Real-time stereo vision on the VisionServer Framework for robot guidance

This paper introduces the VisionServer Framework, a computer vision application framework for development within a visual environment. Its functionality is demonstrated by documenting the design and theory behind a real-time stereo vision system for robot guidance. VisionServer implements the required run-time tasks of stereo image capture, image rectification, stereo correspondence and data visualisation as function blocks which can be visually manipulated within its sequence editor. In order to run in real-time, VisionServer implements stereo correspondence on a graphics processing unit (GPU). Additionally, an end-user GUI can be created within the framework, making it easy to control system settings. The framework is extendable through custom end-user written function blocks. These features promote rapid application development, code reuse, and application sharing. Results show that real-time stereo on the GPU is cost-effective and beneficial for interactive fine-tuning of a stereo algorithm. Moreover, real-time processing is important for time-critical computer vision applications.