Inferring information from video is desirable since video
data is becoming more accessible due to improvements in recording
and storage. Machine learning techniques are also rapidly maturing
with new methods being applied to datasets in different fields. I
have been keenly interested in gathering knowledge from past
experience or test results and applying this knowledge to new
conditions. An approach that I outline in my thesis utilizes a new methodology
for reducing the dimension of large variable number data streams
and developing a probability density function from known conditions
to make estimates on new data.

An example related to fluid mechanics utilizes high speed video
of a flow boiling system was recorded at different heat flux
levels.[i] Nine videos were used in
the analysis corresponding to five values of heat flux which are
shown below at selected points in time. The larger values of
heat flux are shown to have a larger number of bubbles due to the
higher amount of energy being transferred to the working fluid. It
is apparent that the change in the number of bubbles is not
linear.

To demonstrate the result of the method, two videos at heat
fluxes of 80 ^{kW}/m^{2}, 120
^{kW}/m^{2}, 140 ^{kW}/m^{2}, and
160 ^{kW}/m^{2}, were used as the known inputs and
the heat flux was predicted for the 100
^{kW}/m^{2}video. Each image in the 340 frames of
the video corresponds to 141 x 400 pixels resulting in 19,176,000
elements or variables to evaluate per movie. The 19 million
variables are reduced to 64 variables which are embedded into a two
dimensional diffusion space (Shown in the Figure below). These two
element vectors are used to predict the heat flux. Thus, the two
diffusion coordinates of the eight videos corresponding to four
known heat fluxes are used to construct a probability distribution
for the remaining video which is shown in the Figure below.

The resulting mean of the probability density function closely
predicts the actual heat flux applied. Thus, one can potentially
use video of known operating conditions to predict new states.

[i] C. ESTRADA-PEREZ, E.
DOMINGUEZ-ONTIVEROS, H. AHN, N. AMINI, and Y. HASSAN, "PTV
Experiments of Subcooled Boiling Flow through a Rectangular
Channel," in16thInternational Conference on Nuclear Engineering,
Orlando, Florida, 2008.