hii i m pari ..plzzz show the pdf changes in my locality....i want help ....plzzz...
Posts: 391
Threads: 0
Joined: Jul 2016
Introduction
Automatically detecting and segmenting a moving object
from a monocular video is useful in many applications
like video editing, video summarization, video coding, visual
surveillance, human computer interaction, etc.
. Learning moving object cues
We consider moving objects in a video as some compact
regions with different apparent motion from the background.
Specifically, if a region is moving in a certain
frames, we consider it a moving object throughout the
video. This is reasonable in practice. For example in a
video, a boy walks for a while and stops to swing his hands.
It is meaningful to treat his whole body as a moving object
instead of his hands only in the late frames.
Based on the moving object definition, we consider motion
an important cue to identify it. If a pixel/region has
significant different apparent motion from the background,
it mostly likely belongs to a moving object.
. Motion cues
We assume that the background is dominant in the scene.
Based on this assumption, motion cues are defined as the
discrepancy between the local motion and the global background
motion. Estimating the global motion in a video
has a rich literature of possible solutions [19]. Because
the global motion between consecutive frames is small, we
model it using a homography [19]. We use a SIFT [14]
feature-based method to estimate the homography since it
is robust for processing low-quality videos. Specifically, we
extract SIFT features from each frame, establish feature correspondence
between neighboring frames, and estimate the
homography using the RANSAC [6] algorithm.
Moving object segmentation
Given the learned color and locality cues of the moving
objects, we could extend the MRF model in Section 2.3 by
adding the color and locality cues into the likelihood term
in Equation 4 to estimate the full moving objects. A further
step could be to extend the MRF model from one key
frame to the whole video to achieve temporal consistency.
However, since motion cues are sparse and incomplete, and
the locality cues are available only for key frames, the color
cues will be dominant. This can cause false moving object
detection when the background has regions with similar
color. We solve this problem by propagating the locality
cues from each key frame to others. We develop the following
method to propagate the locality cues and segment the
moving objects simultaneously