Ginger/trunk/RTCSA_SS/03Design.tex

\section{System Architecture}
\label{sec:design}

We designed a thermal-box to collect the data. It has four Grideye sensors on the
corners of a 10 cm square and a Lepton 3 at the central. Figure~\ref{fig:method} shows
the system of our method. It consists four parts. The first part is to fuse multiple
data from Grideye sensors into a low-resolution image, since the resolution of a
single Grideye sensor too low to make a decision. The
second part, we train the SRCNN model with fused Grideye image
as low-resolution and downscaled Lepton 3 image as high-resolution image.
The third part, we use the
Super-resolution image to train a neural network model for recognizing current pose
is lay on back or lay on side. The last part, to reduce the noise and effect cause by
the residual heat on bed after turning over. We
remove the noise by median filter, and determine the current pose according to
the trend of the possibility from recognition network.

\begin{figure}[htb]
	\begin{center}
		\includegraphics[width=1\linewidth]{figures/method.pdf}
		\caption{Illustration of Proposed Method.}
		\label{fig:method}
	\end{center}
\end{figure}

\subsection{Grideye Data Fusion}

On the thermal-box, there are four Grideye sensors. At the beginning, we let
the thermal-box faces to an empty bed and records the background temperature.
All the following frames will subtract this background temperature. After that,
we resize four $8 \times 8$ Grideye images to $64 \times 64$ by bilinear
interpolation and then merge them dependence on the distance between thermal-box and
bed, distance between sensors and the FOV of Grideye sensor. In our case, $D_B$ is
150 cm, and $D_s$ is 10 cm.

\begin{enumerate}
\item $D_b$ is the distance between bed and thermal-box.
\item $D_s$ is the width of sensor square also the distance between adjacent sensors.
\item $F$ is the FOV of Grideye sensor which is about 60 degree.
\item $Overlap = 64 - 64 \times (\frac{D_s}{2 \times D_b \times tan(\frac{F}{2})})$
\end{enumerate}

\subsection{Turning Over Determination}

We train a SRCNN model by the fused Grideye image and downscaled Lepton 3 image,
and use it to enhance all following Grideye frames to SR frames. We labeled some SR frames
into two categories, lay on back and lay on side. Since the input
data is very small, we use a neural network consist one 2D convolution layer, one
2D max pooling, one flatten and one densely-connected layer. The possibility of
output has a very large various just after turn over because the model cannot
distinguish the residual heat on bed and the person as Figure~\ref{fig:residual_heat} shown. This
situation will slowly disappear after one or two minutes.

To determination the pose, first we use a median filter with a window size of five
to filter out the noise. Then, find the curve hull line of the upper bound and
lower bound of the data. Finally, calculate the middle line of upper bound and
lower bound, and regrad it as the trend of the pose changing. Figure~\ref{fig:trend}
shows the filitered data and these lines.

We divide every data into 10 second time windows. If the middle line of the time window
is at the top one fifth, or the trend is going up, it is a lay on back. If it is at the
bottom one fifth, or the trend is going down, it is a lay on side. If there are three
continuously same poses, and different from the last turning over, it will be count as
another turning over.

\begin{figure}[ht]
  \centering
  \minipage{0.3\columnwidth}
    \includegraphics[width=\linewidth]{figures/Lepton_residual_heat.png}
    \caption{Residual heat on bed.}
    \label{fig:residual_heat}
  \endminipage
  \minipage{0.65\columnwidth}
    \includegraphics[width=\linewidth]{figures/MinMax.pdf}
    \caption{Trend of pose.}
    \label{fig:trend}
  \endminipage
\end{figure}