Dense stereo based on the uniqueness constraint

The paper presents the matching core of a stereo algorithm suitable for real-time applications. Unlike most area-based algorithms, the proposed approach relies on a single matching phase (i.e. do not include the check for left-right consistency). Unreliable disparity measurements are primarly detected on the basis of the violation of the uniqueness constraint. In order to further improve the reliability of the matches we enforce additional contraints based on the behaviour of the error function that can be verified at a very small computational cost. Experimental results show that the proposed approach provides reliable disparity measurements and that it is significantly fast.