Cryptography Reference
In-Depth Information
part in the previous frame. The foreground pixels in the current frame can
be identified if their color changes significantly after the camera motion. In
addition, information about local motion can also be derived if one subtracts
the estimated camera motion from the original motion vector field. If several
foreground pixels within a macroblock survive the erosion process, the local
motion corresponding to the macroblock can be regarded as robust. Otherwise,
we ignore the local motion, because it probably arises from the background.
After the camera motion has been compensated for, we can extract the
local motion fields from the bitstreams. Each field represents the movement of
a macroblock between two consecutive anchor frames. However, a field is not
usually continuous in the spatio-temporal domain, since motion vectors always
start at regular places, but could end anywhere. Thus, when we try to make
the necessary connections between two local motion fields, each macroblock in
the current frame may overlap with four macroblocks in the previous frame,
as shown in Fig. 9.10. In other words, it is possible that several candidate local
motions located in the previous frame could be connected to a local motion
in the current frame. Therefore, we use the following three criteria to choose
the most reliable candidate for the current local motion: (1) the consistency
of motion direction, (2) the color distribution, and (3) the overlapping area
between macroblocks in the previous and the current frames. Suppose LM
Cur
is a local motion and C denotes the set of candidates of LM
Cur
in the previous
anchor frame. The most reliable candidate, LM
ancestor
, is chosen according to
the following rule:
LM
ancestor
= arg min
m∈C
(θ
m
,αφ
m
,βϕ
m
),
(9.14)
where θ, φ,andϕ denote the angle formed by LM
Cur
and a possible candidate
m, the color histogram difference between macroblocks, and the overlapping
area between macroblocks, respectively. Furthermore, all the parameters are
normalized so that all values fall within the range [0,1], and α and β are
weighted values that balance the influence of the three factors. Once the most
reliable candidate of LM
Cur
has been determined, we link the local motions
acquired at different time spots to form a single motion flow for each mac-
roblock. Since a large moving object may cover several similar motion flows,
we remove redundant flows and preserve one or more representative motion
flows from all the similar motion flows. Initially, we select a motion flow, a,
that has the longest duration in a shot. Suppose a starts from time t
i
and
ends at time t
j
, and letBbe a set of motion flows whose start and end
times are both within the duration of t
i
and t
j
. We remove a motion flow b,
belonging toB, if it satisfies the following condition:
(a
t
−a
t
s
)−(b
t
−b
t
s
)
t∈[t
s
,t
e
]
<ǫ,
(9.15)
t
e
−t
s
where t
s
and t
e
are respectively the start and end times of motion flow b; a
t
and b
t
denote the spatial position of a and b at time t respectively; and a
t
s
