Scale Invariant Feature Transform (SIFT) Parametric Optimization Using Taguchi Design of Experiments

Traditional SIFT methods require a priori of object knowledge in order to complete accurate feature matching. The usual means is via trained databases of objects. In order to be able to get the pose of an object, accurate object recognition is required. Without accurate object recognition, detection can occur but no information about 3-D location will be available. The goal of this work is to improve object recognition using SIFT by optimizing algorithm parameters with respect to the mean angle between matched points (µAMP) found within a scene via multiple images, which can then be used to determine the object pose. Good parameters are needed so that the SIFT algorithm is able to control which matches are accepted and rejected. If keypoint information about an object is wrongly accepted, pose estimation is inaccurate and manipulation capabilities in a 3-D work space will be inaccurate. Using optimized SIFT parameter values results in a 19% improvement of µAMP-Optimal in comparison to µAMP-Experimental.