Resolving stereo matching errors due to repetitive structures using model information

This study regards the problem of incorrect stereo matches due to the occurrence of repetitive structures in the scene. In stereo vision, repetitive structures may lead to “phantom objects” in front of or behind the true scene which cause severe problems in scenarios involving mobile robot navigation or human-robot interaction. To alleviate this problem, we propose a model-based method which is independent of the specific stereo algorithm used. The basic idea is the feedback of application dependent model information into the correspondence analysis procedure without loosing the ability to reconstruct scene parts not described by the model. The employed scene models may either consist of a single plane or (for modelling more complex objects) of several connected planes. An FFT-based detection stage allows the extraction of scene parts displaying repetitive structures and yields the orientation of the model plane, while the plane distance is inferred with a robust optimisation technique based on a model-free stereo analysis. Alternatively, motion-based segmentation can be applied. Our experimental evaluation performed on manually labelled real-world scenes showing objects in front of repetitive structures shows that the proposed method reduces the fraction of false correspondences on the repetitive structures by factors of up to 30 while only moderately decreasing the fraction of 3D points correctly assigned to the object.