Automatic Semantic Video Object Segmentation
No Thumbnail Available
Date
2007-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Content-based video processing is one of the methods considered to meet the demands of
newly emerging multimedia applications. For content-based processing, video has to be
segmented into meaningful objects – semantic video objects. Many applications require
automatic segmentation (AS) of semantic video objects. However, AS is very challenging
task.
In this thesis, an effective AS system is proposed by examining, selecting and combining
efficient and simplified techniques that are justified with theoretical analysis. The proposed
system is developed and tested with the following three cases depending on whether there
exists camera motion or not in video sequences, and whether there is an initial background
reference frame. Case-1 is for video sequences with no camera motion and with initial
background reference frame, Case-2 is for video sequences where there is no camera motion
and no initial background reference frame, and Case-3 is for video sequences with camera
motion.
Change detection is used as the main step in each of the cases to detect semantic objects and
to produce object mask. Different problems in change detection like “uncovered
background”, “global motion of background (GMOB)” and “camera noise” are identified and
solved. Two change detection results are combined to remove the uncovered background
problem. More emphasis is made for the problem of GMOB. A 3-level block-based
hierarchical motion estimation and affine parameter model for frame warping is used to solve
this problem. Camera noise is removed by using model-based change detection.
For post-processing to improve the resulting change detection masks, a new filling-in
technique is proposed. This technique is used to fill open areas inside object regions with
uniform intensity. To improve the boundary of segmentation masks, morphological open
close operations are used. The final semantic video objects are obtained by superimposing the
resulting mask over the original frame.
x
Test results show that the system effectively identified and segmented the semantic objects.
Subjective evaluation of results for the three cases showed that among the window sizes used
in change detection, 5x5 and 7x7 produced better and comparable results in terms of visual
quality and boundary smoothness. These results are obtained after applying one pair of open
close operation in Case-1 and two pairs in Case-2 and Case-3.
Based on subjective comparison of results with other systems, 80% of observations for the
results of Case-1 and 100% for Case-2 reported more pleasing results with smooth boundary.
Keywords: content-based video processing, automatic segmentation, semantic video objects,
change detection, motion estimation, post-processing.
Description
Keywords
Content-based video processing, Automatic segmentation, Semantic video objects, Change detection, Motion estimation, Post-processing