J. Sole, Y. Huang, and J. Llach (USA)
SIFT, RANSAC, mosaic, foreground extraction.
This work deals with the problem of background
foreground segmentation in video scenes. We propose an
approach that makes use of feature extraction and match
ing, robust estimation, background mosaic generation, mo
saic back-projection, and segmentation. We make contri
butions in each of these steps.
SIFT features are extracted from the video and then,
features between consecutive frames are matched. There
are mismatches as well as errors in the feature location. To
surmount this, we use a robust approach that employs a
modified version of the RANSAC algorithm and weighted
total least squares. Knowing the global motion allows cre
ating an initial foreground segmentation and the generation
of a mosaic mainly using background data. The moving
object extraction is done by combining a mean shift-based
frame segmentation along with the information given by
the error of the mosaic back-projection into each frame.