\section{Experiments}


The experiments were carried out on all floors (0 to 3) of the faculty building. 
Each floor is about \SI{77}{\meter} x \SI{55}{\meter} in size, with a ceiling height of \SI{3}{\meter}.
To resemble real-world conditions, the evaluation took place during an in-house exhibition. 
Thus, many places were crowded and Wi-Fi signals attenuated.
As can be seen in fig. \ref{fig:paths} we arranged 4 distinct walks, covering different distances, critical sections and uncertain decisions leading to multimodalities. 
The ground truth is measured by recording a timestamp at marked spots on the walking route.
When passing a marker, the pedestrian clicked a button on the smartphone application. 
Between two consecutive points, a constant movement speed is assumed.
Thus, the ground truth might not be \SI{100}{\percent} accurate, but fair enough for error measurements.
The approximation error is then calculated by comparing the interpolated ground truth position with the current estimation.
Especially in the context of smoothing, it is also very interesting to exclude the temporal delay from the error calculations and measure only the positional difference between estimated and ground truth path. 
This gives a statement about the extent to which the smoothed path superficially improves compared to the filtered one. 

All walks start with a uniform distribution (random position and heading) as prior for $\mStateVec_0$. 
To allow the system to stabilize its initial state, the first few estimations are omitted from error calculations.
Even though, the error during the following few seconds is expected to be much higher than the error when starting with a well known initial position and heading.

The measurements were recorded using a Motorola Nexus 6 and a Samsung Galaxy S5.
As the Galaxy's \docWIFI{} can not be limited to the \SI{2.4}{\giga\hertz} band only, its scans take much longer than those of the Nexus: \SI{3500}{\milli\second} vs. \SI{600}{\milli\second}.
Additionally, the Galaxy's barometer sensor provides fare more inaccurate and less frequent readings than the Nexus does. 
This results in a better localisation using the Nexus smartphone.	
The computation for both filtering and smoothing was done offline using the aforementioned \mbox{CONDENSATION} algorithm and multinomal (cumulative) resampling. 
%However, the filter itself would be fast enough to run on the smartphone itself ($ \approx \SI{100}{\milli\second} $ per transition, single-core Intel\textsuperscript{\textregistered} Atom{\texttrademark} C2750).
%The computational times of the different smoothing algorithm will be discussed later. 
Unless explicitly stated, the state was estimated using the weighted arithmetic mean of the particles. 

As mentioned earlier, the position of all \docAP{}s (about 5 per floor) is known beforehand. 
Due to legal terms, we are not allowed to depict their positions and therefore omit this information within the figures.
Additionally, we used five \docIBeacon{}s for slight enhancements in some areas.
The empirically chosen values for \docWIFI{} were $P_{0_{\text{wifi}}} = \SI{-46}{\dBm}, \mPLE_{\text{wifi}} = \SI{2.7}{}, \mWAF_{\text{wifi}} = \SI{8}{\dB}$, and $\mPLE_{\text{ib}} = \SI{1.5}{}$ for the \docIBeacon{}s, respectively.
The system was tested by omitting any time-consuming calibration processes for those values.
We therefore expect the localisation process to perform generally worse compared to standard fingerprinting methods \cite{Ville09}. 
However, incorporating prior knowledge and smoothing will often compensate for those poorly chosen system parameters.

For the  filtering we used $\sigma_\text{wifi} = \sigma_\text{ib} = 8.0$ as uncertainties, both growing with each measurement's age.
While the pressure change was assumed to be \SI{0.105}{$\frac{\text{\hpa}}{\text{\meter}}$}, all other barometer-parameters are determined automatically (see \ref{sec:sensBaro}). 
The step size $\mStepSize$ for the transition was configured to be \SI{70}{\centimeter} with an allowed derivation of \SI{10}{\percent}. The heading deviation in \refeq{eq:transSimple}, \refeq{eq:transShortestPath} and \refeq{eq:transMultiPath} was \SI{25}{\degree}.
Edges departing from the pedestrian's destination are downvoted using $\mUsePath = 0.9$.

% all paths we evaluated
\begin{figure}
	\input{gfx/eval/paths}
	\caption{The four paths that were part of the evaluation.
	Starting positions are marked with black circles.
	For a better visualisation they were slightly shifted to avoid overlapping.}
	%\commentByFrank{font war korrekt, aber die groesse war zu gross im vgl. zu den anderen}
	\label{fig:paths}
\end{figure}


%kurz zeigen das activity recognition was bringt
By adding the activity recognition to the system of \cite{Ebner-16}, we are able to further improve the overall localisation results. 
The approximation error decreases by an average of \SI{66666}{\centimeter} for all 4 paths on 10 MC runs.
Due to this additional knowledge, the state transition samples mostly depending upon the current activity and therefore limits the possibility of false floor changes.
Fig. \ref{dd} shows the recognized activities for path 4 using the Nexus 6.
Despite a short misdetection caused by faulty pressure readings, the recognition can be considered to be very robust and accurate.

%Fixed Interval Smoothing
At first, both FBS and BS are compared in context of fixed-interval smoothing.
As a reminder, fixed-interval smoother are using all observations until time $T$ and therefore run offline, after the filtering procedure is finished.
Thus, we calculate only the positional error between estimation and ground truth, since timely information are negligible. 

In contrast to BS, the FBS is not able to improve the results using the weighted arithmetic mean for estimating the current position. 
As illustrated in fig. \ref{}, the estimated position for filtering and FBS is identical although the weights are highly different. 
To address this problem, we are calculating the average state using the \SI{50}{\percent} best weighted particles.
For evaluating the FBS this estimation method was applied to filtering and smoothing likewise.
We deployed 10 MC runs using \SI{2500}{} particles for approximation. 
%The resulting positional error between estimated and ground truth along path 4 can be seen in fig. \ref{}. 
Now, the positional error along all 4 paths could be improved from \SI{}{} to \SI{}{} for the galaxy and from \SI{}{} to \SI{}{} for the nexus.
%Similar outcomes can be observed by adding a resampling step at the end of every smoothing iteration. 

However, BS still outperforms the FBS by an average of \SI{}{} on all 4 paths using the same number of particles and \SI{500}{} sample realisations. 
A visual example comparing both smoothing methods on path 4 is illustrated in fig. \ref{}.
Another advantage of BS over FBS, is the ability to still improve the filtering results even while reducing the number of particles radical. 
For example \SI{50}{} particles and \SI{25}{} sample realisations are providing reliable estimations similar to above experiments, though the risk of losing track is higher.

Despite the very good outcomes provided by both interval smoother, there are some rare situations in which smoothing does not improve the filtered estimation or even improves the visual path. 
For example fig. \ref{} depicts such a situation for path 3 using BS and measurements provided by the Galaxy S5.
This shows that the smoothing results are of course highly depend upon the filtering performance.

At next, we discuss the advantages and disadvantages of utilizing FBS and BS as fixed-lag smoother. 
Compared to fixed-interval smoothing, timely errors are now of higher importance due to an interest on real-time localization. 
Especially interesting in this context are small lags $\tau < 10$ considering filter updates near \SI{500}{\milli\second}.

%as seen fbs war im fixed interval schon nicht so gut, im lag ist sein einfluss vernachlässigbar. optische und error technische verbesserung sind kaum vorhanden. lediglich eine verbesserung von  deshalb konzentrieren wir uns bei der diskussion auf den BS. trotzdem verschlechtert sich das ergebniss aber auch nicht. die verbesserung ist nur nicht so signifikant wie bei bs

Fig. \ref{} illustrates the estimation results for path 4 using \SI{2500}{particles}, \SI{50}{sample realisations} for BS and a fixed-lag $\tau = 5$. 

fixed-lag reduces the error about... however, as seen in fig. \ref{} ist der bloße error nicht unbedingt ausschlaggebend für die verbesserung. fast immer liefert smoothing pfade die realistischer sind, aber die error erhöhen. 

%conclusion der experimente
bei weniger partikeln bringt fixed-lag und fixed-interval smoothing im verhältnis sogar mehr! weil es da mehr zum "aufräumen" gibt. trotzdem hängt die performane natürlich stark vom vorwärtsschritt ab und man sollte nicht all zu wenige waehlen und lieber auf nummer sicher gehen. bs ist im vergleich zu fbs ein gutes stück besser in unserem fall. das hängt auch stark mit dem bereits sehr guten filtering schritt zusammen. man könnte aber trotzdem schlussfolgern das bs besser für indoor ist.

%fbs ist hier mies und liegt direkt über dem filter? oder ein fbs mit anderer estimaton und den filter nicht anzeigen?! das ist doch quatsch... wennn ich den filter net anzeige. notfalls einfach den fbs nicht nehmen. sondern sagen das er net taugt und rauswerfen.

%Smoothing mit großen lag kann die zeitliche information schwer halten. das liegt hauptsächlich daran, das im smoothing nur die relativen positionsinfos genutzt werden. das wi-fi wird nicht beachtet und deswegen können absolute justierungen der position (sprünge) nur sehr schlecht abgefedert werden. 


%fixed lag smoothing verbessert den pfad ein wenig und vor allem bügelt unrealistische sprunge aus. daher wird der zeitlich verzug gut herrausgerechnet. 

%beispiel multimodalität

%ein paar worte zur berechenzeit.

%Evaluation:
%\begin{itemize}
%	\item Filter ist immer der gleiche mit MultiPathPrediction und Importance Factors
%	\item FBS Interval mit 500 und 2500 Partikeln auf 4 Pfaden mit SimpleSmoothingTrans
%	\item BS Interval mit 500 zu 100 und 2500 zu 500 Partikeln auf auf 4 Pfaden mit SimpleSmoothingTrans
%	\item FBS Lag = 5 mit 500 und 2500 Partikeln auf 4 Pfaden mit SimpleSmoothingTrans
%	\item BS Lag = 5 mit 500 zu 100 und 2500 zu 500 Partikeln auf auf 4 Pfaden mit SimpleSmoothingTrans
%	\item BS Lag zu Error Plot. Lag von 0 bis 100, wie verhält sich der Error. Am besten auf Pfad 4 mit SimpleSmoothingTrans.
%\end{itemize}