| The complex planning and repositioning of mechanical stages in conventional range sensing of free-form objects, led to development of range sensors permitted continuous free motion. However, to determine the position of the moving range-sensor head, these devices employ mechanical linkages with position sensors, and optical or magnetic tracking devices, which limit the working volume, impede the motion of the range-sensor head, or restrict the working environment.;This thesis presents a method of continuous and unconstrained range-sensing without requiring measurement of the pose, the position and orientation, of either the range-sensor head or object by peripheral sensors, tracking systems, or positioning devices. A continuous sequence of range images which have little sensor movement between views is used to minimize viewpoint planning and to facilitate registration of adjacent views. Views consists of only six range profiles each, to permit rapid acquisition of a view in a single camera image, and to minimize redundant data, while providing sufficient overlap between views for registration.;The registration of adjacent range views, is first explored using correspondence methods based on surface curvature maps and localized spin images. Iterative methods which compute temporary correspondences between views, and three transformation-parameter optimization techniques are also investigated, based on closest-point distances between views. Global registration of a sequence of views to a single reference frame is computed based on inter-view transformations between all adjacent view pairs.;In tests on pairs of synthetic range views of a face, optimization methods yielded the lowest registration errors. Errors were generally an order of magnitude lower than those found in conventional range sensing. For global registration of a sequence of 49 synthetic views of a face, cancellation of inter-view errors over many views, was evident by the error accumulation to lower than peak values over a sequence, and a nearly constant level of error over the latter half of the sequence. Mean absolute global registration errors were still low and acceptable for many surface modeling applications.;The method eliminates the restrictions of other methods which utilize position sensors, and is applicable when the range-sensor head, the object, or both, have continuous unknown motion. |
