An Investigation of Real-time Matching and Pose Estimation

An analysis is presented of different methods of recognition based on image template matching. They are studied with respect to robustness, speed and precision of the estimated relative object position and orientation (pose). In this investigation, the matching problem was modeled in terms of PRISM sign bits of the Laplacian of Gaussian (SLOG) image correlation [Nishihara84 ] and a practical, but sufficiently general, recognition scenario. In the analysis, the error tolerances on the 3D pose estimate given the scenario are calculated. This is followed by an experimental study of the stability and false positive rate of SLOG matches under varying pose. The analysis is completed by an evaluation of each recognition approach with respect to the desired pose error tolerances and the experimentally determined performance of SLOG matching. Accelerated SLOG-based template matching, relational matching and image encoding all show promise as real-time matching modules, especially in the context..