Rotation clustering for robot vision

摘要

A robot vision system is presented which estimates the orientation and position of an arbitrary known object. The technique relies on signatures approximating the spatial distributions of edges which are generated from observed images and a CAD model of the target under examination. The signature based 3-dimensional vision system generates a map of cross-correlations between training signatures and single image signatures on a domain of rotations. In this dissertation, we treat the signature correlations as a relative indicator of the a posteriori distribution of the target rotation given the correlations. Since the high correlations are subject to noise in the amplitudes of the peaks, a maximum a posteriori (MAP) estimate of rotation can generate incorrect rotation estimates of the target. Instead, we perform a localized MAP estimate of the rotations which estimates the rotation of the target given the actual rotation is in its local neighborhood. The localized map estimates are obtained through the clustering of rotations. To cluster rotations we first examine the theory of rotations and define a useful algebra, norm, and metric on the space of rotations. Next we analyze the problem of averaging rotations through a minimum angle error metric and by examining the performance of several popular techniques. Finally we examine several clustering techniques modified for their use with rotations. In particular, using a weighted version of our rotation averaging techniques, we derive modified K-means and fuzzy c-means clustering algorithms which work on the non-linear space of rotations. The clustering techniques are applied to the signature based robot vision system. We demonstrate a significant improvement in the estimation of rotation by using the modified fuzzy c-means clustering. In addition we extend the system to operate over several images by reapplying the fuzzy clustering to the set of likely rotations to find an average of a consistently good set of rotation results. Experiments show that the multi-image clustering technique out-performs previous techniques of 3-D target pose estimation with only a small increase in computational overhead.