Dynamic integration of camera motion, multiresolution image acquisition and coarse-to-fine surface reconstruction from stereo

This thesis is concerned with the problem of surface reconstruction from stereo images for large scenes having large depth ranges. An active stereo approach is used in which the scene is systematically scanned, and image acquisition and surface reconstruction are integrated.The thesis particularly investigates the situation in which the scanning of an object has been completed and the surface reconstruction process must resume by fixating another object. The approach is based on acquiring coarse structural information in the vicinity of the next fixation point during the current fixation, and refining this information while reconfiguring the cameras to fixate the next point. It involves the following steps. First, a new fixation point is selected from among the nonfixated, low resolution scene parts of current fixation. Second, a reconfiguration of cameras is initiated for refixation. As reconfiguration progresses, the images of the new fixation point gradually deblur, and the accuracy of the stereo estimate for the point improves allowing the cameras to be aimed at it with increasing precision. The improved stereo estimate guides the selection of focus settings for the cameras. Finally, within the fixated parts of the scene focus-based depth estimates are obtained at a gird whose density is determined by the local surface slope.The selection of new target points on objects other than the one currently being scanned is cast as an objective function to be minimized. The criteria favor visual targets that are near the current fixation point, both laterally and in depth. The cost of the objective function is measured in terms of the time taken to mechanically refixate the cameras and to stereo analyze the images acquired during such refixation.The refixation phase is modeled as an error reduction process iterating over the steps of surface reconstruction and camera readjustments. Integration of the cues of stereo, focus and vergence is performed to reach the state of fixation. The depth estimates from these cues are combined at the end of the fixation process to form a statistically more accurate estimate for the fixation point. Experimental results based on a particular implementation of the approach are presented. (Abstract shortened with permission of author.)U of I OnlyETDs are only available to UIUC Users without author permissio