Image interpretation using multisensor integration

The dissertation studies two aspects of multi-sensor image interpretation: the low-level integration of segmentation maps and the high-level integration of interpretation knowledge. AIMS, the system developed during the research, uses registered laser radar and thermal images for scene interpretation. Its objective is to detect and recognize man-made objects in outdoor scenes. Information from four sensing modalities (range, intensity, velocity, and thermal) are integrated to improve both image segmentation and interpretation using real data.; An algorithm that integrates region-based segmentation maps and edge maps is developed. The algorithm operates independently of image sources and specific segmentation techniques. Any number of mixed types of segmentation maps and user-supplied weightings on these maps are allowed. The initial estimation stage uses the maximum likelihood criterion and enforces contour connectivity using edge adjacency information. The resultant edge contours are smoothed iteratively by minimizing a potential function. Finally, regions are merged to guarantee that each region is large and compact. The initial estimation and the contour smoothing are both spatially controlled by the channel resolution width for multi-scale processing.; The low-level attributes of image segments (regions) are computed using the integrated segmentation map. Forward chaining is used in a bottom-up fashion to derive object-level interpretations from data collected by the low-level processing modules. Segments are grouped into objects and objects are classified into pre-defined categories. Non-symbolic processing tasks are dispatched to a concurrent service manager. Therefore, vision tasks with different characteristics are executed using different software tools and methodologies.; The newly developed low-level integration algorithm allows multiple segmentation modules to work in parallel as front-ends. Consequently, the need is reduced to develop a single powerful segmentation technique for a wide range of sensors and image contents. The same technique can be extended to work with multiple shape-from-X schemes by providing the mechanism needed to integrate multiple cues. Multiple information sources, integrated image segmentation, and the hypotheses integration technique enable AIMS to tolerate intermediate errors by verifying image segmentation and interpretation incrementally. The techniques developed in the research are applicable to other related domains, such as autonomous navigation, remote sensing, and object recognition.