Available online at www.sciencedirect.

com

Procedia Social and Behavioral Sciences 20 (2011) 506–513

14th EWGT & 26th MEC & 1st RH

Situational Analysis in Real-Time Traffic Systems

Abstract

The prediction of traffic situations is a vital issue in modern Intelligent Transport Systems (ITS). Particularly the real time
detection and proper assessment of incidents may save live and may contribute to keep the transport network available. However,
the influence factors of traffic are subject to multi changes; the influence factors are varying traffic demand, weather, seasons,
etc. This allows only a very poor performance in real time assessment of traffic situations. This article focuses on the fact that the
possible traffic patterns – depicted as time series – vary only very little on each site, representing specific traffic situations or
“normal time series”.
That the use of intelligent dedicated digital signal processing systems and communication media is in a position to improve the
requirement of Real-Time traffic situational analysis in an efficient and effective way. Digital Finite Impulse Response (FIR)
filters are used to analyse sensor measurement data in a way that allows an instantaneous assessment about the actual traffic
situation and the detection of abnormal behaviour. A FIR filter cascade is used, each filter represents a specific normal time
series like working day, weekend day, rain, winter, etc. The representation of each normal time series is achieved by discrete
transformation of the normal time series in order to generate its frequency spectrum and by designing a filter structure with a
corresponding frequency response.
The deployment of the FIR filters can be in Traffic Control Centres with a vast amount of computational power as well as in the
controller cabinets of local sensors on the basis of Digital Signal Processors.
© 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and/or peer-review under responsibility of the Organizing Committee.
Keywords: ITS systems; FIR filters;traffic situations;Traffic Control Centres

1. Introduction

The general task of TMSC is to keep transport infrastructure available, to reduce the negative effects of traffic
and to maintain or raise the safety level on the roads – all with the means of ITS. This covers the gathering of traffic
raw data, the managing of the traffic flow, and the detecting of incidents. The quality of service of a TMSC can be
expressed in its operating cost in relation to the time to reaction or in relation to the rate of disturbance reduction.
The treatment of incidents requires real time data and a situation analysis in real time or near real time on the tactical

* Corresponding author. Tel.: +12 6341568; fax: +12-6341568.

1877–0428 © 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and/or peer-review under responsibility of the Organizing Committee
doi:10.1016/j.sbspro.2011.08.057
 Dietrich Leihs and Andrzej Adamski / Procedia Social and Behavioral Sciences 20 (2011) 506–513 507

level as well as on the strategic level (Adamski, 1999a, 1999b, 2002). However, incidents have to be judged in the
context of the general (traffic) situation in order to determine their severity and relevance.
A typical way of determining the actual traffic state is supplied by the German standard MARZ and FGSV 358
(German Federal Agency for road affairs, 1999). According to MARZ, the traffic situation is classified into 4 levels
according to defined mean velocities and the local traffic density on the various lanes. A more sensitive approach to
determine the traffic situation is delivered by FGSV 358 (German Road and Transportation Research Association,
1992), calculating the actual traffic status on a respective road section on the basis of the fundamental diagram
(speed-volume-scatter). The most recent approach of traffic situation awareness in traffic control is the definition of
standard time series (Adamski, 1999). Basically, traffic volume measurements are combined with an event calendar;
though the correlation procedures are inflexible and do not reflect the “similarity” of traffic patterns. This procedure
aims at the traffic situation evaluation, it is hardly possible to qualify and quantify specific situation as they occur.
The scientific community shows a more ambitious approach towards the traffic situation identification. Neural
network models are used for predicting the traffic situation under adverse conditions (Huang & Ran, 2003, Eun-Mi,
Jai-Hoon and Won-Sik, 2007). Since the neural network has the potential of solving nonlinear problems it can tackle
an input-output mapping, it may be used for predicting complex traffic situations (Haykin, 1998, Innamaa, 2000,
Val, Larrodé and Gea, 2005). Another approach within the scientific community is fuzzy clustering (Nowotny,
Asamer, Din and Rouzbeh, 2003, Tao, 2002, Adamski & Habdank-Wojewódzki, 2005). The Fuzzy clusterisation
features are its global range and the circumstance that the clusters are fuzzy, so the decision is not very crisp, thus
small changes of the data will not highly influence the result. However the algorithm is complicated and a high
computational power in case of optimization is needed. On the other hand, the hierarchical fuzzy subtractive
clusterisation is efficient and demands lower computational power demand but needs a manual stop. The
combination of the two methods is unsupervised fuzzy clustering with multi-centre clusters (Tao, 2002), supplying
proper results. By not looking for multiple centres – which is the case for single centre traffic patterns – the
clusterisator is working based on minimizing the measure of the differences between the data points and the cluster
centres. By fuzzy clustering a time series of a whole day, one centre for the whole period will be found. This allows
that the centre points of whole time series can be compared, which is particularly of little help when determining the
actual morning peak. A more tricky approach is to look for fragments of time series that are similar to other
fragments within the same time series and within other fragments of other time series. As a result of the
clusterisation process, a “toolbox” of typical scenarios in the time domain is created.

Figure 1 Fuzzy clustered centres in a time series of traffic quantities over a period of seven days. The cluster length is 1 hour (i.e. 4 measuring
values). Source: (Leihs & Adamski, 2007)
508 Dietrich Leihs and Andrzej Adamski / Procedia Social and Behavioral Sciences 20 (2011) 506–513

2. Real Time Filtering of Traffic Time Series

2.1. Real Time Classification of Traffic Patterns

The traffic data measured will always reflect all factors influencing the respective traffic situation. Hence by
collecting and suitable classifying of historical traffic data, it is possible to estimate the influence factors. The result
is a set of behavioural patterns of a measuring site, allowing the cause analysis and setting of parameters of the
influence factors. The result of this process is Normal Time Series, which allow a long range modelling or
prediction of the traffic situation. During real time operation it is extraordinarily important to know if the actual
traffic situation is according to the long range model or not in order to set proper measures according to the
application. Digital filters are suitable for this task. The filter characteristics of the filters reflect typical behavioural
patterns of a specific traffic situation. The filter characteristics are gained from the behavioural patterns via z-
transformation and subsequent filter design methods. The actual measuring values are filtered with these filters; the
filter response is a clear, qualitative and quantitative indication for the correlation of the actual measuring values
with the respective behavioural pattern.

2.2. Creating a Classification Filter

In order to determine the filter parameters, every normal time series has to undergo a discrete transformation like
z-transformation order to determine its frequency spectrum. Any digital filters can be designed by distributing poles
and zeros in the complex z-plane. The transfer function of a FIR (Finite Impulse Response) filter is (Antoniou,
2000)
M

N M
–
i 1
z  zi
H ( z) H0z N

– z  zi
i 1 (1)
where the numerator polynomial represents the zeros and the denominator polynomial the poles of the transfer
function. The stability condition requires that the roots of a denominator polynomial are inside the unit circle of the
complex plane. It is important to emphasize that H(ejZ) is the value of the z-transform H(z) on the unit circle
(0≤ Z≥2S) (one turn around the unit circle). From ej2Sk = cos (2Sk)+j sin (2Sk) = 1 results that H(ejZ) is periodic with
j Y 2Sk
the period 2S. Hence H (e ) H (e jZ ) .
FIR filters have one pole in the origin of the z-plane; in order to achieve any filter characteristic, zeros can be put
along the unit circle. With other words, the filter characteristic of a FIR filter can be described adequately by the
number and position of zeros. This property eases the use of FIR filters.
The transfer function (i.e. frequency characteristic) will be transformed into a simplified model under solely by
using zeros. It has to be mentioned that a zero appears as a local minimum in the frequency spectrum, respectively
on the unit circle in the complex plain. The closer the zero is to the unit circle, the lower is the numeric value of the
local minimum in the spectrum. If the zero is far from the unit circle, the local minimum is rather “flat”.
The next step covers the design of the digital filter by means of the detected zeros. The Gaussian fundamental
theorem of the algebra allows transforming the zeros, and by expanding the resulting polynomial factors the filter
coefficients are produced. Now it is possible to create the filter with the structure of figure 2, where the coefficients
a1...n are integrated into the multiplier stage.
 Dietrich Leihs and Andrzej Adamski / Procedia Social and Behavioral Sciences 20 (2011) 506–513 509

Figure 2. Digital filter synthesized from the zeros of the transformed normal time series

The digital filter is fed with actual numeric measuring values in real time. These values are filtered and the result
of the filtering process can be taken from the output. As it is not possible to determine the filter characteristics in
terms of classical filter design (high pass, low pass, band pass, etc.), a classification stage is necessary. The principle
of the classification stage is a time synchronous comparison of two filter responses: the one to the input signal and
the other one to the normal time series. The more the signals are similar, the lower is the difference. The difference
vector will undergo a root mean square (RMS) calculation in order to get a single value result. The numeric value of
this single value result is the closer to zero the more the input signal is corresponding to the normal time series. The
scheme of the classification stage is shown in figure 3.

Figure 3. Set-up of the chain for real time classification of time series

After each normal time series is integrated in an individual digital filter, all filters are integrated into the
classification stage shown in figure 3, and all classification stages are connected in parallel to a filter cascade:

Figure 4.Parallel filter cascade

At the input, real time (RT) numeric measuring values are fed into the cascade. The output of the cascade is a
single numeric value for each filter of the cascade. The very filter with the filter characteristic corresponding best to
the input time series produces the number closest to zero.
As each digital filter within the cascade represents a specific normal time series, and as each normal time series
represents a specific behavioural pattern, the output of the parallel filter cascade indicates the behavioural pattern
that corresponds most to the actual behaviour. Furthermore, each behavioural pattern might have been undergone a
cause analysis and parameterisation of its influence factors, so it is possible to determine if the actual behaviour is as
usual or if it is deviating from the normal behaviour. Though, normal behaviour is a situation sensitive issue;
something which is normal on working days might not be normal on weekends; something which is normal during
510 Dietrich Leihs and Andrzej Adamski / Procedia Social and Behavioral Sciences 20 (2011) 506–513

dry weather might not be normal during rain or snow. Each digital filter represents the behavioural pattern of a
specific traffic situation.

3. Performance

The performance of this approach was validated by means of traffic measurements on a motorway (AIT Austrian
Institute of Technology, 2002). Between July 10th and July 29th, 2002, the vehicle quantities and velocities were
determined on the first two of three lanes with an averaging interval of 10 minutes. The time profile of the
measurement is shown in figure 5.

Figure 5. Time profile of velocities and volumes of the test site; source: (AIT Austrian Institute of Technology, 2002)

For further treatment, two normal time series were generated – for simplification one for all Wednesdays and one
for all Sundays – by averaging the respective daily time series. Those two normal time series were used to create a
filter structure as described above. These normal time series are shown in figure 6. The result of the discrete
transformation is shown in figure 7.
In order to investigate the quality performance of the filter, a look is granted into the classification stage (see
figure 3): the result of the subtraction of the normal time series filter and the input filter that was performed in
Matlab is shown figure 8. As an input signal, the measuring values of single days were fed into the input; the normal
time series filter was reflecting either the Wednesday normal time series or the Sunday normal time series.
In the top image the Wednesday filter is used, the filter response to every week day is subtracted from the filter
response to the Wednesday normal time series. In the bottom image the Sunday filter is used, the filter response to
every week day is subtracted from the filter response to the Sunday normal time series. The optimal auto correlation
would be a horizontal line at 0.
In the bottom image (Sunday filter), the shape of the Sunday behaves as assumed and correlates the best with the
ideal horizontal line at 0. So this scenario proved to work. In the top image (Wednesday filter) however it can be
observed that all working days follow more or less to the ideal correlation, the Friday changes from a positive to a
negative trend due to a congestion in the afternoon. Saturday and Sunday are deviating in a strong way. However, an
outstanding performance of the Wednesday cannot be spotted, as the normal time series of Mondays...Thursdays are
alike.
This leads to the claim for high quality time series. The approach selected here – i.e. the “arbitrary” clusterisation
according to the name of the weekday – corresponds to real life. In particular the necessity to characterize the
measuring values of a huge number of detector sites makes it impossible to classify the curves manually according
 Dietrich Leihs and Andrzej Adamski / Procedia Social and Behavioral Sciences 20 (2011) 506–513 511

to the similarity of their pattern. For that reason the quality of the resulting normal time series can be influenced by
considering the shape of the curve at the clusterisation.
Another point of discussion is the length of the normal time series. In this case the classification criteria “week
day” lead to time series with the length of 24 hours. During a whole day though, the characteristics of road traffic
might be subject of significant change.

Figure 6. Normal time series for traffic volume for Wednesdays and Sundays

Figure 7. Result of the z-transformation of the two normal time series for Wednesdays and Sundays as a spectrum
512 Dietrich Leihs and Andrzej Adamski / Procedia Social and Behavioral Sciences 20 (2011) 506–513

Figure 8. Difference of the filter response to single weekdays and to the normal time series; top: Filter = Wednesday; difference of all week days
to the Wednesday normal time series bottom: Filter = Sunday; difference of all week days to the Sunday normal time series
 Dietrich Leihs and Andrzej Adamski / Procedia Social and Behavioral Sciences 20 (2011) 506–513 513

4. Summary and outlook

The selected approach of employing digital filters with filter parameters representing historical traffic behaviour
turns out to be suitable to detect “abnormal” traffic behaviour (incidents) and to assess the relevance and criticality
of those incidents in real time. Due to the use of digital filters, this method is suitable for a low resource ambient like
a DSP as well as for an ambient with a vast computing power.
Hence this method may be a valuable contribution for Traffic Management Centres in order to increase the
capabilities in terms of real time behaviour, in terms of cost efficiency of the employed measures and last but not
least in terms of the effectiveness of the centres. Particularly regarding the effectiveness, there is a lack of suitable
methods that are qualifying and quantifying traffic disturbances in real time. In this way, the deployment of digital
filters allows the situation sensitive assessment of traffic situations which is the precondition for the evaluation if a
certain situation is in a certain respect abnormal and/or critical for the road network. Due to the scalability of digital
filters, a large quantity of road elements that are equipped with measuring equipment (as is the case with floating car
data or with entire motorway networks) can be efficiently supervised in an intelligent way. Due to the fact that
digital filters can be deployed on a DSP, this is even possible on the level of a single sensor.

References

Adamski A. (1999 a). Integrated Transportation Systems. Modeling and management in transportation, EWG T, 1/21-34.
Adamski A. (1999 b), Integrated Pro-ecological Traffic Planning and Management Approach. Proc. 6th Meeting of the EWGT, Chalmers.
Adamski A. (2002), Intelligent Integrated Transportation Systems. 9th Meeting of the EWGT, “Intermodality, Sustainability and ITS”, Bari 2002.
Adamski A., Habdank-Wojewódzki S. (2005). Traffic congestion and incident detector realized by fuzzy discrete dynamic system, The Archives
of Transport PAN, XVII(2):pp.5-13.
AIT Austrian Institute of Technology. 2002.
Antoniou A. (2000), Digital Filters: Analysis, Design, and Applications, McGraw-Hill Science/Engineering/Math; 2nd edition.
Eun-Mi L., Jai-Hoon K., Won-Sik Y. (2007). Traffic Speed Prediction Under Weekday, Time, and Neighboring Links’ Speed: Back Propagation
Neural Network Approach, Springer Berlin / Heidelberg.
German Federal Agency for road affairs (1999). Bulletin for the equipment of traffic management centres and sub centres.
German Road and Transportation Research Association (1992), Hints for the traffic flow analysis, incident detection and traffic flow prognosis in
traffic management of non urban roads.
Haykin S. (1998). Neural Networks, A Comprehensive Foundation. Prentice Hall.
Huang S.H., Ran B. (2003). An Application of Neural Network on Traffic Speed Prediction Under Adverse Weather Condition, Transport
Research Board Annual Meeting 2003, Paper Nr. 000915.
Innamaa S. (2000). Short-Term Prediction of Traffic Situation using MLP-Neural Networks, Proceeding Nr. 2075, ITS World Congress Torino.
Leihs D., Adamski A. (2007). Real Time Situational Awareness – A New Approach In Classifying The Actual Traffic Behaviour, ITS ILS 2007.
Nowotny B, Asamer J, Din K, Rouzbeh K. (2003). Classification of traffic data time series by cluster analysis, artificial neural networks and
ANOVA, 10th World Congress on Intelligent Transport Systems, Madrid, Spain.
Tao C.W. (2002). Unsupervised fuzzy clustering with multi-center clusters, Elsevier Science B.V., Fuzzy Sets and Systems 128 (2002) 305–322.
Val S., Larrodé E., Gea Á. (2005). Industrial Vehicles Park Capacity Sizing By Means Of Artificial Neural Networks, Proceeding Nr. 43, EWGT,
Poznan.