Page 134 - Proceedings of the 2018 ITU Kaleidoscope

2018 ITU Kaleidoscope Academic Conference

the quantized motion vector orientation of the ith block. In
Algorithm the computation of influence weight, it is known that only a
Input: K − Set of blocks in the frame pair of blocks are considered, that is w reflects only the

Output: M −Motion Descriptor Map influence of the i block on block j. To compute the motion
th
M is set to zero at the beginning of each frame influence vector of the jth block within a frame, assumption
for all i in K, is based on all other blocks that potentially affect the

ℎ = max( ).size motion of block j.

for all j K where ≠ ji, = ∑ (7)
Compute - Euclidean Distance between ( )

block i and j Where j ∈ {1, 2, . . ., MN}, i denotes the quantized
if ≤ ℎ , orientation index of a block, which is used as a component

− Angle btw block i and j index of block-j and denotes the computed influence

k =⌊ / 45⌋ weight. The computed motion influence weight of the block

= exp(- / ) of every frame is added up to form a motion descriptor map.

= + The motion descriptor map is clustered into many clusters
based on the influence weights using k-means clustering.
end if The result of the k-means clustering provides a behavioral
end for pattern which has influence weights in many clusters.
end for
3.4. Nearest Neighbor Search
The movement of the object is highly influenced by
neighboring blocks relative to a selective block which In this module, the influence weights computed from the
concludes that neighboring blocks will have high influence previous modules are framed as motion descriptor map and
over the distant blocks. For every individual block searched for nearest cluster distance in the trained system.
structure, a distinct parameter called threshold distance is If the motion descriptor is near to center of any cluster of
being computed by multiplying magnitude of each block the behavior pattern then it is a normal block. If the
with total block size of an overall frame. Threshold th is distance between the computed motion descriptor and

computed as, closest cluster center must be lesser than threshold of

= ( ). (2) acceptance, then the block is considered normal. If the

distance is greater than threshold then the block in the

where, b is the magnitude of block b in the direction k. It frame is considered abnormal. Minimum distance, md of

is calculated by multiplying the size of single block with deviation of the computed motion descriptor is calculated
magnitude of particle pointing to similar direction. For as,

every block in a frame, Euclidean distance between every =∀ min ( ( )) (8)
other block and the corresponding angle of deviation The block is considered abnormal if md is greater than the
between blocks are computed. The flag variable f is threshold of acceptance.
computed as,
0 , > ℎ
= (3)
1 , ℎ
Let θ be the angle of deviation between block i and j.
= ⌊ / 45⌋ (4)
Now, influence weight of blocki on blockj, w is computed

as,

= . (− / ) (6)

Influence weight, w of blocks is calculated for every

frame in the video and added with influence weight of
previous blocks called Motion Descriptor.

3.3. Motion Descriptor Pattern Clustering Figure 4 - Visualization of detection of abnormal block in
nearest neighbor search
After computing the influence weights of all blocks (w ), a

motion influence map clustering is generated which
significantly represents the motion patterns within a frame.
Each component of the motion influence vector represents

– 118 –