Information selection and fusion in vision systems


Handling the enormous amounts of data produced by data-intensive imaging systems, such as multi-camera surveillance systems and microscopes, is technically challenging.While image and video compression help to manage the data volumes, they do not address the basic problem of information overflow. In this PhD we tackle the problem in a more drastic way. We select information of interest to a specific vision task, and discard the rest. We also combine data from different sources into a single output product, which presents the information of interest to end users in a suitable, summarized format. We treat two types of vision systems. The first type is conventional light microscopes. During this PhD, we have exploited for the first time the potential of the curvelet transform for image fusion for depth-of-field extension, allowing us to combine the advantages of multi-resolution image analysis for image fusion with increased directional sensitivity. As a result, the proposed technique clearly outperforms state-of-the-art methods, both on real microscopy data and on artificially generated images.The second type is camera networks with overlapping fields of view. To enable joint processing in such networks, inter-camera communication is essential. Because of infrastructure costs, power consumption for wireless transmission, etc., transmitting high-bandwidth video streams between cameras should be avoided. Fortunately, recently designed ‘smart cameras’, which have on-board processing and communication hardware, allow distributing the required image processing over the cameras. This permits compactly representing useful information from each camera. We focus on representing information for people localization and observation, which are important tools for statistical analysis of room usage, quick localization of people in case of building fires, etc.To further save bandwidth, we select which cameras should be involved in a vision task and transmit observations only from the selected cameras. We provide an information-theoretically founded framework for general purpose camera selection based on the Dempster-Shafer theory of evidence. Applied to tracking, it allows tracking people using a dynamic selection of as little as three cameras with the same accuracy as when using up to ten cameras.