Page 133 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

module where each frame is divided into uniform blocks
and optical flow magnitude of each blocks are calculated in
different directions. Then, maximum magnitude in the
block is taken as threshold value and the magnitude of the
particles in other blocks pointing towards the block is taken
for computing the influence weights. Similarly for every
frame in the video, influence weight is calculated and
stored in Motion Descriptor map. The generated motion
descriptor map is used for training the system by k-means
clustering algorithm which clusters the motion descriptor
map based on the influence weights as data points. For each
block separate clusters are formed.

3.1. Magnitude Vector Computation

The video dataset is split into individual frames for
processing them in a sequential manner. Each frame is
analyzed specifically for every moving pixel and it is being
identified as individual particle. The optical flow is more
effective in calculating the displacement between all frames
which takes brightness as its pattern. Each particle has
individual magnitude and directions which are all the Figure 3(a) - Visualization of a block with optical flow
properties by which a particle differs from an idle pixel. movements inside the block
Individual frame in video dataset is divided into uniform
blocks. Block classification is more efficient than object
classification as it is faster and computation of foreground
extraction is not needed. In addition, object classification
needs trajectory extraction which is a complex task but it is
as effective as block identification. In block detection,
unwanted object movements may collide as noise and it is a
challenging task to be eliminated. Our current approach is
based on a block size of 8x6 which is fixed for maintaining
an accurate performance state.

For every block magnitude vector of size 8 is assigned and

magnitude of particles moving in a 45 pace are added
together to represent the magnitude of the block as eight
octant in terms of direction. For instance, magnitude of

particles with direction of motion between 0 and 45 are

added together and the resultant value occupies b where b

represents magnitude vector of the ith block. An individual
block in frame with particles with different magnitude and
direction is depicted in Figure 3(a) and 3(b) showing the
particles pointing to similar direction.
Figure 3(b) - Classification of Optical Flow movements
The particles pointing towards similar directions are inside a block based on direction of movement
grouped together and total sum of magnitude of all particles
in each group is calculated and stored in an array. 3.2. Motion Descriptor Computation
Maximum value in array is selected as the threshold

magnitude of the block. The magnitude of the block b in The main feature of the proposed motion influence map is

direction k is calculated as, that it effectively reflects the motion characteristics of the
( ). /
movement speed, movement direction, corresponding

= (1) magnitude and size of the objects or subjects, and their

. / interactions within a frame sequence. Each frame is
separated into several block structures for computing
where p represents direction of motion of a particle and individual motion features of every block with its all other

p represents magnitude of the particle. neighboring blocks. Two major factors which influence the

movement of objects are: i. magnitude, and ii. direction
angle.

– 117 –