Discrete object detection and motion registration based on a data management approach

When scientific data sets can be interpreted visually they are typically managed as pictures and consequently stored as large collections of bitmaps. Valuable information contained in images is often not exploited, however simply because the data is not processed further. Common reasons for this are that access to information in image collections is notoriously difficult and that interesting applications often depend on supplementary data with incompatible formats. If such data sets are treated as higher-dimensional point data instead of byte streams and managed with a suitable multidimensional file structure, then it is possible to transform "fuzzy" objects into n-dimensional solids. Several benefits result: content based access becomes possible, the potential for data compression without loss of relevant information exists and additional information can readily be incorporated simply by increasing the file structure's dimensionality. This paper describes how this approach has been successfully applied to detect and track storm cells in weather radar and various satellite images. The key is to parametrize the data for efficient access based on several different image attributes.