Subspace position measurement in the presence of occlusion

An eigenspace-based method for measuring camera and/or object position in the presence of occlusions is described. Following the formation of a global eigenspace during a training phase, input images are divided into subsections for which a local set of projections is stored. Local occlusions in each subsection are detected by eigenspace reconstruction error. Only unoccluded subsections are then recombined dynamically to implement the nearest-neighbor search of the manifold and to avoid ambiguity in the subspace. Experimental results are described to demonstrate this technique for the measurement of the position of two metallic parts in the presence of significant occlusion.