Comparison and evaluation of similarity measures for vergence angle estimation

This paper presents a comparison of various similarity measures, developed for the real-time control of the vergence angle in an active vision robot head. The vergence angle can be estimated using several classical difference or correlation measures. These methods are studied comparatively for various images sizes. The Zero-mean Normalized Cross Correlation (ZNCC) metric is proved to outperform the other methods. The results also show that we can sufficiently control the vergence mechanism, using images even 256 times smaller than the original (i.e. 40x30 pixels), in less than 1ms. We evaluated these results using two methods. Firstly, we extracted the disparity maps of the stereo pair to test whether the disparity value is zeroed when we reach the correct vergence angle. Secondly, we calculated the Mean Square Error (MSE) and the Normalized MSE (NMSE) of the correlation index between the sub-sampled and the initial images.