Autonomous Input Management for Human Interaction-Oriented Systems Design Michal Podpora Opole University of Technology Faculty of Electrical Engineering, Automatic Control and Informatics ul. Sosnkowskiego 31, 45-272 Opole, Poland Email: michal.podpora@gmail.com Aleksandra Kawala-Janik, Mary Kiernan University of Greenwich, School of Computing and Mathematical Sciences, Old Royal Naval College Park Row, SE10 9 LS London, UK Email: {a.d.kawala-janik, m.kiernan}@greenwich.ac.uk Abstract—In this paper evaluation of a policy-based algorithm for video inputs switching is presented. The term ’data quality’ is not trivial for Human-Machine Interaction systems, yet a simple and efficient algorithm is needed for choosing the most valuable video source. This becomes particularly important for systems that support functional decomposition of image pro- cessing algorithm, which are designed for non-optimal working environment. In this paper an autonomous input management system is proposed, which consists of a data quality evaluation algorithm and a simple decision algorithm. Keywords – Human-Machine Interaction (HMI), Machine Vision, Decision Systems, Distributed Systems, Autonomic Systems I. I NTRODUCTION A N INCREASING number of computer and robotic systems designed for interaction with humans are using computer vision as its basic information source. Advances in vision– and sound-based interaction are fundamental for developing efficient and fluent interactions. The most recent implementations and research conducted have proven that it is possible to implement an ample vision– and speech-based man-machine interaction system, however the tests were run in a laboratory environments only. In real life applications the issue is much more challenging, especially in case, where various environments and conditions are taken into account. However, there are some mechanisms in biology, which are able to handle both data acquisition and information processing in conditions exceeding the capabilities of the very acquisition subsystems. As an example – human can see in a dark room without the daylight, human can read and understand words even if the majority of letters are illegible. Whilst the goal of the interaction is not only entertainment, computer systems and robots are also supposed to be able to support and to protect humans. To do that, these systems would have to contain a more efficient acquisition and information processing subsystems than humans. It is not difficult to find future applications for systems that expand the acquisition possibilities of human and interact using speech and/or vision: fire rescue teams support robotic systems, traffic collision and pedestrian avoidance systems, etc. . . II. POLICY-BASED I NPUT SWITCHING It is theoretically possible to implement all available acquisition systems and to process all the input information (vision, thermal video, night vision, etc. . . ), but in real-life applications it is not efficient because the data streams are often too massive and require high computing power for real-time processing and inferring, so the implementation in an embedded system would be impossible. Transmission of multiple video streams to a remote workstation/server is also not always possible due to the limited bandwidth [1]. However, it is possible to choose one input at a time and transmit it (or process it locally) [2], if the system was able to judge the value of the data input quality of its acquisition subsystems. The quality evaluation could be performed on the basis of static threshold values (e.g. any measure of image noise or edges quality), but in that case it would be just a simple condition, with no deeper meaning. The Policy-Based Input Switching (PIS) becomes especially useful when there is a necessity for adjusting/changing the policy for choosing an input depending on circumstances or environment. For example, the visual input of smoke has a relatively low level of noise and a satisfying brightness level, while the input is completely useless and should be switched to thermography (or at least infrared vision). The proposed PIS conception is not just a set of rules and threshold values, but an entire framework offering wrapper functionality (similarly to [3], [4], [5]) for additional modularity and as a foundation for autonomous policy reloading/changing. A semi-autonomous mobile robot, with PIS implemented, designed for supporting a fire team, would be able to have its policy changed for operating in a particular environment and would be able to change its policy autonomously depending on any transient environment parameters (e.g. brightness, opacity, visibility, obstacles,