Stretch-correlation as a real-time alternative to feature-based stereo matching algorithms

摘要

We have analysed the requirements for a robust stereo vision algorithm for use in typical industrial applications. For such applications the views obtained in the two cameras have large differences in visual appearance due to the orientation difference between the two cameras and the close proximity of illumination sources. We have concluded that for this category of problem, feature-based methods should be more robust than conventional, area-based approaches, and this conclusion appears to be borne out in the published literature. However, correlation-based approaches are more suited to efficient implementation on available hardware. The technique which we have developed, called Stretch-Correlation, is based on the cross-correlation of warped image blocks which have been preprocessed to maximize the useful information content. Our new method models the severe warping effects encountered in difficult stereo problems and effectively relaxes the front-o-parallel constraint which is normally imposed in area-based disparity calculation. This algorithm imposes effectively most of the local constraints present in feature-based algorithms, and can be efficiently implemented on available hardware.