REAL-TIME URBAN TRAFFIC STATE-ESTIMATION AND PREDICTION
DESIGN OF A DATA-FUSION & LINK-NEIGHBOURHOOD FRAMEWORK Author: drs. ir. Luuk de Vries The use of information tech-
ronment can therefore make many more decisions (degrees of
nology (IT) in traffic systems
freedom) as opposed to on a freeway, where only lane changes
is becoming a hot topic wit-
and the occasional merging/splitting at ramps are possible deci-
hin the traffic research com-
sion points.
munity. It is within this area within civil engineering that IT
BIG DATA
and traffic, blend together to
Regarding the traffic information and traffic data collection, the
create intelligent transporta-
common practice is to rely solely on a network of roadside
tion systems (ITS). Benefits of
sensors responsible for generating data. These most commonly
employing ITS are ample and can range from: increased safety,
equipped sensors are inductive loop detectors (ILD), which – if
improved operational performance, enhanced mobility, environ-
dually equipped – give information about all three main macro-
mental benefits to boosts of productivity leading to economic
scopic traffic flow variables (flow, occupancy and speed) at fixed
and employment growth. One of the practical outcomes can for
positions in the road network. Inductive loop detector data
example be that all actors of the transportation systems are allo-
(ILDD) comes with two major drawbacks, firstly it is prone to
wed to enlighten themselves with information and make better
errors and secondly the broader spatial representativeness of
informed decisions. This type of utilization of ITS leans on a key
measured traffic flow variables is questionable. As in urban net-
ingredient: a robust, complete and accurate picture of the traffic
works the coverage rate of ILDD is generally low, the measure-
state in the network.This picture is generated in a process called
ments from a limited ILD sample, does not suffice to provide the
traffic state estimation. It generally goes hand in hand with the
complete traffic information needed.
process of traffic state prediction, which generates the future
Radio-frequency identification (RFID) or Bluetooth identifica-
pictures of the traffic state.
tion are alternatives used to obtain individual travel times based on vehicle identification and re-identification. Yet come at high
AN URBAN ENVIRONMENT In my research towards urban traffic state estimation and prediction, two challenges are overcome. Firstly there is a mismatch between the amount of data and the needed traffic flow variables. A traffic network is in practice never fully covered by traffic information sources, thus requiring techniques to extrapolate and utilize the traffic data that is available. The second challenge is a result of choosing an urban environment as subject of study in this research. An urban environment which as opposed to a freeway only network, comes with a higher complexity due to e.g. lower traffic volumes, lower speed limits, more variability in velocities, a higher density of intersections, traffic signals, roundabouts, priority-junctions and dynamic interactions between other modes of transport. A vehicle in an urban envi-
Issue 26-3 1 CONCEPTUEEL
FIG. 1 The generalized traffic control cycle in which state-emission and state-prediction are key components. Source: Van Lint, J. (2015)
FIG. 2 A big difference between a freeway-only network (left) and an urban environment (right) are the degrees of freedom of a vehicle. Source: Google Street View (left. I-90 near Sioux Falls, right. Center of Sioux Falls)
cost, privacy concerns, low coverage and can only measure travel
of traffic state estimation and prediction is to find a balance
times between set locations. For License plate recognition (LPR)
between sophisticated and complex models on one side and
and other video image techniques, the same disadvantages apply.
smooth, fast, general applicable models on the other side, to
A more promising and state-of-the-art data source is Floating
make valid estimations and forecasts given the data available.
Car Data (FCD). FCD consists of reported vehicle positions,
The naïve categorization represents models in which only traffic
direction of driving and velocities for timestamps with a pre-
data is used from which direct relations are calculated. Examples
defined temporal spacing from dedicated vehicle probes. These
are instantaneous travel time or historical averaging models.
vehicles are equipped with a form of GPS and a communication
The advantage of these models is the favourable low compu-
link for transferring this data. FCD has the advantage opposed to
tational complexity and easy implementation. The downside
ILDD to be able to determine a representative mean speed for
is that because of the lack of traffic theory, results are usually
a whole road segment. The downsides of FCD are the wobbly
illogical and inaccurate. The parametric categorization repre-
representativeness of FCD due to the penetration rate, reso-
sents models in which the principle of the Lighthill–Whitham–
lution and accuracy as a result of mapping issues due to tall
Richards model are applied. Classical examples are: Newell’s sim-
buildings and complex networks within an urban environment.
plified kinematic wave model and cell transmission models. The
More recently, a futuristic data source is compiled by mining of
advantage of these models is that they implement logical real
traffic jam reactions through online social sensors. These social
world traffic theory, with the disadvantage of requiring vast cali-
media sensors (SMS) gather data by crawling through posts on
bration of parameters. Additionally due to the subject of study
regular social media (e.g. Twitter) or on more specialized traffic
being an urban network where the traffic flow fundamentals of
apps (e.g. Waze). This offers a fast and low cost way to under-
for example flow conservation might not be applicable, accu-
stand what is happening in the physical world, although the noisy
racy is negatively affected. The non-parametric categorization
nature of the data makes quality still lacking.
represents the traffic models in which relations in traffic data
To make the most of all available traffic data, the method or algo-
are considered, but no traffic flow parameters are estimated.
rithm used for both traffic state estimation and traffic state pre-
Examples of these models are mostly based on simple regres-
diction, calls for fusion of heterogeneous data sources to maxi-
sion. They have in common that while their complexity is low
mize the utility of the available information. The question rises
and therefore they can easily be run in real-time speed, though
on how to deal with this ever-growing amount of big data and
their accuracy is generally fairly low.
how to integrate different sources of traffic data as to achieve full potential. But then, the usefulness of this vast amount of big
HYBRID TRAFFIC MODELS
data will still depend on the quality of the recorded data, used
Theoretically more suitable for urban traffic state estimation
extrapolation technique and the used traffic model.
and prediction are hybrid model types, which take elements FIG. 3 The four types of mathematic traffic (flow) models
CLASSICAL TRAFFIC MODELS Generally the vast amount of traffic estimation and predictions models used in literature, can be fitted into four categories. There are naïve models, parametric models, non-parametric models and a hybrid blend of two or more of these categories. The general consensus is that the key difficulty within the area
CONCEPTUEEL Issue 26-3
2
FIG. 4 Traffic state estimation framework. Refined expansion of Tao (2012)
from non-parametric, parametric and naĂŻve methods to output
first part of the research a microsimulation software package
more accurate estimations and predictions. The most famous
(Paramics-Discovery) is used to generate a 100% accurate and
example is Kalman Filtering, which again assumes all traffic flow
100% covered ground truth for the classical Sioux Falls network
fundamentals to hold. Of all the other hybrid models considered
used in literature. This ground truth designed remains unchan-
in this research, the hybrid black-box approaches seem to be
ged throughout. The second part presents the extension of the
theoretically, practically and intuitively, the best suitable in the
ground truth framework with the NLM for traffic state estima-
urban environment chosen as subject of this study. Within these
tion. Additionally the steps of performance assessment, evalu-
lines Morita (2011) and Esaway (2012) separately developed two
ation and synthesis are included to complete the design cycle.
interesting frameworks which consider the use of patterns in
With the ground truth available (for comparative purposes), the
historical traffic data and allow the current traffic state of links
estimations are assessed on accuracy and correlation leading to
to be used as indicators for the traffic state on neighbouring
the designing of the best performing NLM variant for both traffic
links.
state-estimation and state-prediction.
RESEARCH METHODOLOGY
FIG. 5 Map of the designed Sioux-Falls study area within Paramicsdiscovery
The goal of my research is to design a performing traffic state estimation and traffic state prediction framework, which by utilizing both floating car - and inductive loop detector- data, delivers real-time and future link -velocities, -densities and -flows within an urban traffic environment. Used as a starting point for this design are the previously methods of Morita (2011) and Esaway & Sayed (2012), from which this newly developed neighbourhood link method (NLM) is created. Answered is the question on how to design a neighbour link framework which delivers both a traffic state estimation and state prediction of all relevant traffic flow variables within an urban network. Additionally both the performance and accuracy of the traffic states outputted by this NLM framework are assessed. The research method applied is divided into three separate parts: 1) the urban traffic state ground truth, 2) the urban traffic state estimation and 3) the urban traffic state prediction part. In the
Issue 26-3 3 CONCEPTUEEL
FIG. 7 Result example of the traffic state estimation framework for link #40 in the Sioux Falls case study network area
THE NLM FRAMEWORK
ons only drop marginally, but the correlation of flow predictions
The newly developed NLM framework is described with the fol-
lowers by 10% to 15%. For a prediction horizon of 15 minutes
lowing 7 steps. Initially traffic data is stored in a database. Next
the degradation continues as the prediction accuracy decrea-
from this database for each link it is determined which links
ses further and the correlation for density and flow drops well
behave the same and can be considered neighbours based on
below 0,75.
correlation in traffic data. Then newly arrived data is considered upon which an estimation from solely the traffic data of neigh-
FUTURE RESEARCH
bouring links is generated using linear regression. Consequently
This research reveals that the considered NLM framework can
this neighbourhood estimation is fused, weighted on reliability,
yield very reasonable traffic state estimation results in a model-
with the traffic data from the link itself, generating the final traffic
led and simulated environment. Due to the fact it is simple in
state estimation. For the prediction part, an extra time dimen-
essence and algorithmically not very complex, NLM can also be
sion is included, to incorporate the prediction horizon.
easily transferred to a real world scenario. Additionally other traffic data sources can be effortlessly implemented in the pro-
THE RESULTS
cess. There are however areas suitable for further research,
For the die-hard readers interested in the statistical and mathe-
especially as the predictive ability of NLM is currently unsatis-
matical results of this research, this paragraph summarizes the
factory. Advanced bagging of historical traffic data can provide
most important numerical outcomes. The results reveal that the
additional accuracy, as well as a different approach to finding the
NLM framework for estimation at a penetration rate of 5% FCD,
neighbourhood space for each link.The time-lag inherently appa-
is able to estimate on average 60% of the urban links in the
rent in NLM (because the first link that experiences congestion
network within 5 km/h of the ground truth during rush hour
cannot be predicted by its neighbours) has not been overcome.
periods. Density estimations show 80% of all links to be estima-
A plug-in incorporates historical traffic data differently might
ted within 7,5 veh/km/lane and of the flow estimations 60% of
provide a solution here. Additionally further work is needed to
the links can be estimated within 100 veh/h/lane deviations in
improve and test the current implementation into more com-
rush hour. With corresponding correlation values of over 0,91
plex and realistic urban traffic networks as not all traits that
for velocity; 0,93 for density and 0,75 for flow estimations. The
typically describe an urban environment were included in the
NLM framework for prediction shows at 5% FCD and a predic-
used case study (e.g. user-interaction, mode-interaction, a hete-
tion horizon of 5 minutes again promising results, with only a
rogeneous vehicle mix and dynamic traffic lights).
20% drop in accuracy for velocity. However density predictions are up to 50% less accurate, and flow predictions are up to four
Esawey, M. and Sayed, T. (2012). A framework for neighbour
times(!) worse. The correlation of velocity and density predicti-
links travel time estimation in an urban network.
FIG. 6 The seven steps of the proposed NLM-framework
Transportation Planning and Technology, 35:3, 281-301, DOI: 10.1080/03081060.2012.671028
LITERATURE
Morita, T. (2011). High Performance Spatial Interpolation System for Traffic Conditions by Floating Car Data. SEI Technical Review, number 72, April 2011
Tao, S., Manolopoulos,V., Rodriguez, S., Rusu, A. (2012). Real-
Time Urban Traffic State Estimation with A-GPS Mobile
Phones as Probes. Journal of Transportation Technologies, 2012, 2, 22-31
CONCEPTUEEL Issue 26-3
4