\section{Real-World Evaluation}
\label{sec:experiment}

We conduct a real-life uncontrolled experiment using the Kinect to test to the
algorithm.  First we describe our approach to 
data collection.  Second we describe how the data is processed and classified.
Finally, we discuss the results.

\subsection{Dataset} 

The Kinect outputs three primary signals in real-time: a color image stream, a
depth image stream, and microphone output (\fref{fig:hallway}).  For our
purposes, we focus on the depth image stream.  As the Kinect was designed to
interface directly with the Xbox 360, the tools to interact with it on a PC are
limited.  The OpenKinect project released libfreenect~\cite{libfreenect}, a
reverse engineered driver which gives access to the raw depth images of the
Kinect.  This raw data could be used to implement skeleton fitting algorithms,
\eg those of Plagemann~\etal{}~\cite{plagemann:icra10}.  Alternatively,
OpenNI~\cite{openni}, an open framework led by PrimeSense, the company behind
the technology of the Kinect, offers figure detection and skeleton fitting
algorithms on top of raw access to the data streams.  More recently, the Kinect
for Windows SDK~\cite{kinect-sdk} was released, and its skeleton fitting
algorithm operates in real-time without calibration.  

Prior to the release of the Kinect SDK, we experimented with using OpenNI for
skeleton recognition with positive results.  Unfortunately, the skeleton
fitting algorithm of OpenNI requires each individual to strike a specific pose
for calibration, making it more difficult to collect a lot of data.  Upon the
release of the Kinect SDK, we selected it to perform our data collection, given
that it is the state-of-the-art and does not require calibration.

We collect data using the Kinect SDK over a period of a week in a research
laboratory setting.  The Kinect is placed at the tee of a well traversed
hallway.  The view of the Kinect is seen in \fref{fig:hallway}, showing the
color image, the depth image, and the fitted skeleton of a person in a single
frame.  For each frame where a person is detected and a skeleton is fitted we
capture the 3D coordinates of 20 body joints, and the color image.

\begin{figure}[t]
  \begin{center}
    \includegraphics[width=0.99\textwidth]{graphics/hallway.png}
  \end{center}
  \vspace{-\baselineskip}
  \caption{Experiment setting. Color image, depth image, and fitted
  skeleton as captured by the Kinect in a single frame}
  \label{fig:hallway}
\end{figure}

For some frames, one or several joints are out of the frame or are occluded by
another part of the body. In those cases, the coordinates of these joints are
either absent from the frame or present but tagged as \emph{Inferred} by the
Kinect SDK. Inferred means that even though the joint is not visible in the
frame, the skeleton-fitting algorithm attempts to guess the right location.


Ground truth person identification is obtained by manually labelling each run
based on the images captured by the RGB camera of the Kinect. For ease of
labelling, only the runs with people walking toward the camera are kept. These
are the runs where the average distance from the skeleton joints to the camera
is increasing.

\subsection{Experiment design}
\label{sec:experiment-design}

We preprocess the data set to extract \emph{features}
from the raw data.  First, the lengths of 15 body parts are computed from the
joint coordinates. These are distances between two contiguous joints in the
human body. If one of the two joints of a body part is not present or inferred
in a frame, the corresponding body part is reported as absent for the frame.
Second, we reduce the number of features to nine by using the vertical symmetry
of the human body: if two body parts are symmetric about the vertical axis, we
bundle them into one feature by averaging their lengths. If only one of them is
present, we take its value. If neither of them is present, the feature is
reported as missing for the frame. Finally, any frame with a missing feature is
filtered out.  The resulting nine features include the six arm, leg, and pelvis
measurements from \xref{sec:uniqueness}, and three additional measurements:
spine length, shoulder breadth, and head size.  Here we list the nine features as
pairs of joints:
%The resulting nine features are: Head-ShoulderCenter, ShoulderCenter-Shoulder,
%Shoulder-Elbow, Elbow-Wrist, ShoulderCenter-Spine, Spine-HipCenter,
%HipCenter-HipSide, HipSide-Knee, Knee-Ankle.  

\vspace{-1.5\baselineskip}
\begin{table}
\begin{center}
\begin{tabular}{ll}
Head-ShoulderCenter & Spine-HipCenter\\
ShoulderCenter-Shoulder & HipCenter-Hip\\
Shoulder-Elbow & Hip-Knee\\
Elbow-Wrist & Knee-Ankle\\
ShoulderCenter-Spine &\\
\end{tabular}
\end{center}
\end{table}
\vspace{-2.5\baselineskip}

Each detected skeleton also has an ID number which identifies the figure it
maps to from the figure detection stage. When there are consecutive frames with
the same ID, it means that the skeleton-fitting algorithm was able to detect
the skeleton in a contiguous way. This allows us to define the concept of a
\emph{run}: a sequence of frames with the same skeleton ID.

We perform five experiments.  First, we test the performance of skeleton
recognition using traditional 10-fold cross validation, to represent an offline
setting.  Second, we run our algorithms in an online setting by training and
testing the data over time.  Third, we pit skeleton recognition against the
state-of-the-art in face recognition.  Next, we test how our solution performs
when people are walking away from the camera.  Finally, we study what happens
if the noise from the Kinect is reduced.

%\begin{table}
%\begin{center}
%\caption{Data set statistics. The right part of the table shows the
%average numbers for different intervals of $k$, the rank of a person
%in the ordering given by the number of frames}
%\label{tab:dataset}
%\begin{tabular}{|l|r||r|r|r|}
%\hline
%Number of people & 25 & $k\leq 5$ & $5\leq k\leq 20$ & $k\geq 20$\\
%\hline
%Number of frames & 15945 & 1211 & 561 & 291 \\
%\hline
%Number of runs & 244 & 18 & 8 & 4\\
%\hline
%\end{tabular}
%\end{center}
%\end{table}

\begin{figure}[t]
  \begin{center}
    \includegraphics[]{graphics/frames.pdf}
  \end{center}
  \vspace{-1.5\baselineskip}
  \caption{Distribution of the frame ratio of each individual in the
  data set}
  \label{fig:frames}
\end{figure}

\subsection{Offline learning setting}
\label{sec:experiment:offline}

In the first experiment, we study the accuracy of skeleton recognition using
10-fold cross validation.  The data set is partitioned into 10 continuous time
sequences of equal size. For a given recall threshold, the algorithm is trained
on 9 continuous time sequences and trained on the last one. This is repeated
for the 10 possible testing subsamples. Averaging the prediction rate over
these 10 training-testing experiments yields the prediction rate for the chosen
threshold. We test the mixture of Gaussians (MoG) and sequential hypothesis
testing (SHT) models, and find that SHT generally performs better than MoG, and
that accuracy increases as group size decreases.


\fref{fig:offline} shows the precision-recall plot as the threshold varies.
Both algrithms perform better than three times the majority class baseline of
15\% with a recall of 100\% on all people.  Several curves are obtained for
different group sizes: people are ordered based on their frequency of
appearance (\fref{fig:frames}, and all the frames belonging to people beyond a
given rank in this ordering are removed.  The decrease of performance when
increasing the number of people in the data set can be explained by the
overlaps between skeleton profiles due to the noise, as discussed in
Section~\ref{sec:uniqueness}, but also by the very few number of runs available
for the least present people, as seen in \fref{fig:frames}, which does not
permit a proper training of the algorithm.

\begin{figure*}[t]
\begin{center}
\subfloat[Mixture of Gaussians]{
    \includegraphics[]{graphics/offline-nb.pdf}
    \label{fig:offline:nb}
}
\subfloat[Sequential Hypothesis Learning]{
    \includegraphics[]{graphics/offline-sht.pdf}
    \label{fig:offline:sht}
}
  \caption{Results with 10-fold cross validation for the top $n_p$ most present people}
\label{fig:offline}
\end{center}
  \vspace{-1.5\baselineskip}
\end{figure*}

%\begin{figure}[t]
%  \begin{center}
%    \includegraphics[width=0.80\textwidth]{graphics/10fold-naive.pdf}
%  \end{center}
%  \caption{Precision-Recall curve for the mixture of Gaussians model
%  with 10-fold cross validation. The data set is restricted to the top
%  $n$ most present people}
%  \label{fig:mixture}
%\end{figure}

\subsection{Online learning setting}

In the second experiment, we evaluate skeleton recognition in an online
setting.  Even though the previous evaluation is standard, it does not properly
reflect reality. A real-life setting could be as follows. The camera is placed
at the entrance of a building. When a person enters the building, his identity
is detected based on the electronic key system and a new labeled run is added
to the data set. The identification algorithm is then retrained on the
augmented data set, and the newly obtained classifier can be deployed in the
building. 

In this setting, the sequential hypothesis testing (SHT) algorithm is more
suitable than the algorithm used in the previous paragraph, because it
accounts for the fact that a person identity does not change across a
run. The analysis is therefore performed by partitioning the dataset
into 10 subsamples of equal size. For a given threshold, the algorithm
is trained and tested incrementally: trained on the first $k$
subsamples (in the chronological order) and tested on the $(k+1)$-th
subsample. \fref{fig:online} shows the prediction-recall
curve when averaging the prediction rate of the 10 incremental
experiments.

\begin{figure}[t]
%\subfloat[Mixture of Gaussians]{
%    \includegraphics[width=0.49\textwidth]{graphics/online-nb.pdf}
%    \label{fig:online:nb}
%}
%\subfloat[Sequential hypothesis testing]{
\parbox[t]{0.49\linewidth}{
\begin{center}
    \includegraphics[width=0.49\textwidth]{graphics/online-sht.pdf}
\end{center}
    \label{fig:online:sht}
  \vspace{-1.5\baselineskip}
    \caption{Results for the online setting, where $n_p$ is the size of
    the group as in Figure~\ref{fig:offline}}
    \label{fig:online}
}
\parbox[t]{0.49\linewidth}{
  \begin{center}
    \includegraphics[width=0.49\textwidth]{graphics/face.pdf}
  \end{center}
  \vspace{-1.5\baselineskip}
  \caption{Results for face recognition versus skeleton recognition}
  \label{fig:face}
}
\end{figure}

\subsection{Face recognition}

In the third experiment, we compare the performance of skeleton recognition
with the performance of face recognition as given by \textsf{face.com}.  At the
time of writing, this is the best performing face recognition algorithm on the
LFW data set~\footnote{\url{http://vis-www.cs.umass.edu/lfw/results.html}}.
The results show that face recognition has better accuracy than skeleton
recognition, but not by a large margin.

We use the publicly available REST API of \textsf{face.com} to do face
recognition on our data set.  Due to the restrictions of the API, for this
experiment we train on one half of the data and test on the remaining half. For
comparison, MoG algorithm is run with the same training-testing partitioning of
the data set. In this setting, SHT is not relevant for the comparison, because
\textsf{face.com} does not give the possibility to mark a sequence of frames as
belonging to the same run. This additional information would be used by the SHT
algorithm and would thus bias the results in favor of skeleton recognition.
However, this result does not take into account the disparity in the number of
runs which face recognition and skeleton recognition can classify frames, which
we discuss in the next experiment.

\begin{figure}[t]
\parbox[t]{0.49\linewidth}{
  \begin{center}
    \includegraphics[width=0.49\textwidth]{graphics/back.pdf}
  \end{center}
  \vspace{-1.5\baselineskip}
  \caption{Results with people walking away from and toward the camera}
  \label{fig:back}
}
\parbox[t]{0.49\linewidth}{
  \begin{center}
    \includegraphics[width=0.49\textwidth]{graphics/var.pdf}
  \end{center}
  \vspace{-1.5\baselineskip}
  \caption{Results with and without halving the variance of the noise}
  \label{fig:var}
  }
\end{figure}

\subsection{Walking away}

In the next experiment, we include the runs in which people are walking away
from the Kinect that we could positively identify.  The performance of face
recognition ourperforms skeleton recognition the previous setting.  However,
there are many cases where only skeleton recognition is possible. The most
obvious one is when people are walking away from the camera. Coming back to the
raw data collected during the experiment design, we manually label the runs of
people walking away from the camera. In this case, it is harder to get the
ground truth classification and some of runs are dropped because it is not
possible to recognize the person. Apart from that, the data set reduction is
performed exactly as explained in Section~\ref{sec:experiment-design}. Our
results show that we can identify people walking away from the camera just as
well as when they are walking towards the camera.

%\begin{figure}[t]
%  \begin{center}
%    \includegraphics[width=0.80\textwidth]{graphics/back.pdf}
%  \end{center}
%  \caption{Precision-Recall curve for the sequential hypothesis
%  testing algorithm in the online setting with people walking away
%  from and toward the camera. All the people are included}
%  \label{fig:back}
%\end{figure}

\fref{fig:back} compares the curve obtained in \xref{sec:experiment:offline}
with people walking toward the camera, with the curve obtained by running the
same experiment on the data set of runs of people walking away from the camera.
The two curves are sensibly the same. However, one could argue that as the two
data sets are completely disjoint, the SHT algorithm is not learning the same
profile for a person walking toward the camera and for a person walking away
from the camera. The third curve of \fref{fig:back} shows the precision-recall
curve when training and testing on the combined dataset of runs toward and away
from the camera.

\subsection{Reducing the noise} 

For the final experiment, we study what happens when the noise is reduced on
the Kinect.  
%Predicting potential improvements of the prediction rate of our
%algorithm is straightforward. The algorithm relies on 9 features only.
%\xref{sec:uniqueness} shows that 6 of these features alone are sufficient to
%perfectly distinguish two different skeletons at a low noise level. Therefore,
%the only source of classification error in our algorithm is the dispersion of
%the observed limbs' lengths away from the exact measurements.
To simulate a reduction of the noise level, the data set is modified as
follows: we compute the average profile of each person, and for each frame we
divide the empirical variance from the average by 2. Formally, if $\bx$ is an
observation in the 9-dimensional feature space for the person $i$, and if
$\bar{\bx}$ is the average of all the observations available for this person in
the data set, then $\bx$ is replaced by $\bx'$ defined by:
\begin{equation}
  \bx' = \bar{\bx} + \frac{\bx-\bar{\bx}}{\sqrt{2}}
\end{equation}
We believe that a reducing factor of 2 for the noise's variance is
realistic given the relative low resolution of the Kinect's infrared
camera. 

\fref{fig:var} compares the Precision-recall curve of
\fref{fig:offline:sht} to the curve of the same experiment run on
the newly obtained data set.

%\begin{figure}[t]
%  \begin{center}
%    \includegraphics[width=0.49\textwidth]{graphics/var.pdf}
%  \end{center}
%  \vspace{-1.5\baselineskip}
%  \caption{Results with and without halving the variance of the noise}
%  \label{fig:var}
%\end{figure}

%%% Local Variables: 
%%% mode: latex
%%% TeX-master: "kinect"
%%% End: