Using temporal seeding to constrain the disparity search range in stereo matching

Abstract

In a stereo image sequence, finding feature correspondences is normally done for every frame without taking temporal information into account. Reusing previous computations can add valuable information. A temporal seeding technique is developed for reusing computed disparity estimates on features in a stereo image sequence to constrain the disparity search range. Features are detected on a left image and their disparity estimates are computed using a local-matching algorithm. The features are then tracked to a successive left image of the sequence and by using the previously calculated disparity estimates, the disparity search range is constrained. Errors between the local-matching and the temporal seeding algorithms are analysed. Results show that although temporal seeding suffers from error propagation, a decrease in computational time of approximately 20% is obtained when it is applied on 87 frames of a stereo sequence. Furthermore a confidence measure is used to select start-up points for seeding and experiments are conducted on features with different confidences.