Trafﬁc State Estimation on Urban Roads

Using Perception-Enriched Floating Car Data

Moritz Schweppenh

auser

1 a

, Karl Schrab

2 b

, Robert Protzmann

1 c

and Ilja Radusch

Fraunhofer Institute FOKUS, Kaiserin Augusta Allee 31, 10589 Berlin, Germany

TU Berlin / Daimler Center for Automotive IT Innovations, Ernst-Reuter-Platz 7, 10587 Berlin, Germany

Keywords:

Trafﬁc State Estimation, Simulation, Vehicle Perception, Eclipse MOSAIC.

Abstract:

Modern-day navigation systems by developers like Google© and TomTom© require user participation primar-

ily in the form of Floating Car Data (FCD) for accurate Trafﬁc State Estimation (TSE). However, to provide

reliable information, systems rely on large road user participation of at least 5 %, which is only truly available

to the big players. We propose a method to soften the participation requirement by utilizing modern perception

sensors (e.g., radar, lidar, camera) of connected vehicles (CVs) to enrich the FCD set, compensating reduced

data quantity with increased data quality. By using position and speed estimates of surrounding vehicles we

increase the sample size and can additionally collect estimates of segments that are not traversed by CVs. To

validate and assess the proposed system, we utilize Eclipse MOSAIC and conduct a simulation-based test se-

ries on the calibrated large-scale BeST scenario. Initial ﬁndings indicate improved estimation performance on

selected road segments, especially at lower rates of market penetrations. In a network-wide investigation, we

show that travel time estimates of the proposed method are often more accurate than conventional approaches,

while also requiring smaller penetration rates.

1 INTRODUCTION

In a world highly dependent on personal mobility,

trafﬁc congestion is a daily issue, leading to wasted

time, increased emissions, and fuel consumption.

Navigation system developers try to alleviate the traf-

ﬁc strain by providing real-time Trafﬁc State Estima-

tion (TSE) for users, giving them the ability to avoid

trafﬁc build-ups and congested areas. To provide ac-

curate estimations, current-day systems typically rely

on large amounts of data provided by users. This data

is referred to as Floating Car Data (FCD) and is usu-

ally made up of a vehicle’s position, its heading, and

speed. This technology has been established over the

last three decades and nowadays is an essential part of

most navigation applications.

At the same time, automotive sensor technolo-

gies evolved to add richer perception sensors, such

as radar, lidar, and cameras. This trend has been

driven by advanced driver assistance systems and au-

tonomous driving technology, reaching a state where

https://orcid.org/0009-0001-9252-2425

https://orcid.org/0000-0002-5083-595X

https://orcid.org/0000-0002-5531-1936

many new car models are equipped with such sensors.

In related work, Delooz et al.(Delooz et al., 2020)

utilize V2X (vehicle-to-everything) communication

to share the advanced sensor data between vehicles

or servers to enable innovative use cases. This ﬁeld of

research is mainly driven by safety use cases such as

collective perception.

Inspired by these use cases, we want to investigate

the shared sensor output as an input for trafﬁc efﬁ-

ciency applications. Speciﬁcally, we propose a TSE

system that extends FCD packages with information

about surrounding vehicles, including the respective

position, heading, and speed. With this approach, we

aim to further reduce market penetration thresholds of

conventional FCD-based systems, potentially allow-

ing smaller competitors to enter the market.

Similarly, Messelodi et al.(Messelodi et al., 2009)

propose to enrich FCD with data from sensors on the

equipped vehicles. Their approach includes warning

sign recognition as well as a local trafﬁc level esti-

mator, relying on onboard preprocessing. Ruppe et

al.(Ruppe et al., 2012) further propose to augment

FCD with data from surrounding Bluetooth/Wi-Fi en-

abled devices, increasing the amount of collected data

without the use of perception sensors.

Schweppenhäuser, M., Schrab, K., Protzmann, R. and Radusch, I.

Trafﬁc State Estimation on Urban Roads Using Perception-Enriched Floating Car Data.

DOI: 10.5220/0012620500003702

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 10th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2024), pages 99-111

ISBN: 978-989-758-703-0; ISSN: 2184-495X

The work presented in this paper is built on pre-

viously published research, which gave an initial re-

view on sensor modalities and aggregation methods

for mean speed estimation(Schweppenh

auser et al.,

2023). We extend the established system and test

the novel solution using the co-simulation framework

Eclipse MOSAIC(Schrab et al., 2023).

This paper is structured into four main sections.

First, in Section 2 we reiterate fundamental require-

ments in terms of TSE relevant to this paper and ex-

plain how one could include potentially lossy data

from surrounding vehicles. Secondly, we explain how

we approach the implementation and deployment of

the system in a simulated environment in Section 3.

Conducted experiments and compiled results are de-

scribed in Section 4. Finally, with Section 5 we lay

out closing remarks and future research directions.

2 MEAN SPEED ESTIMATION

Estimating trafﬁc ﬂows, speeds, and densities has

long been an important topic in trafﬁc research. In

real-world applications, this task quickly becomes

difﬁcult as only limited measurement options exist,

each coming with different coverage, reliability, and

costs. Historically, static observers such as induc-

tion loop sensors and trafﬁc cameras have been de-

ployed on major road arteries to investigate their traf-

ﬁc state with good accuracy. However, these static

installations are costly which inherently limits their

network coverage. Contrarily, the idea of using mo-

bile sensors in the form of connected vehicles (CV),

has been proposed as early as the 1950s(Wardrop

and Charlesworth, 1954) using the “Moving Observer

Figure 1: Exemplary space-time diagram in an urban sce-

nario, highlighting large deviations in vehicle trajectories.

The dark purple line indicates the calculation of the tempo-

ral mean speed for the marked spatio-temporal interval.

Method”. This method relies on vehicles being able

to perceive vehicles in the opposing lanes, which is

why it was never widely deployed, as required sensors

were not available in the past. Later on, with the larger

availability of mobile connectivity and the universal

adoption of smartphones, the “Floating Car Observer

Method”, i.e., using Floating Car Data, gained popu-

larity as this enabled the possibility for almost every

end user to act as a mobile sensor.

Navigation application developers such as

Google© and TomTom© rely on FCD for Trafﬁc

State Estimation. However, for aspiring competitors,

the FCD method is hardly applicable as large amounts

of data are required, ranging to 5 % to 10 % market

penetration(Ferman et al., 2005) even for estimations

on highways. On urban roads, this matter is further

complicated due to much higher variability in driver

behavior and a more complex road layout.

Independent of the applied sensor technology, the

task of trafﬁc state estimation describes the recon-

struction of different trafﬁc indexes such as ﬂows,

densities, mean speeds, and other metrics built on

the former based on potentially sparse sensor data.

Blokpoel et al.(Blokpoel et al., 2010) describe an un-

ambiguous list of such trafﬁc metrics used for classi-

fying the trafﬁc state. Building on our previous work

we continue focusing on mean speed estimates, as the

speed is one of the core indexes and can be directly

used in fastest-path routing algorithms.

To better understand the role of speeds in the traf-

ﬁc state so-called Space-Time-Diagrams can help (see

Figure 1). We, conventionally, plot the time on the ab-

scissa and the space on the ordinate. Trajectories of

individual vehicles are displayed as purple lines and

the slope of a trajectory indicates a vehicle’s speed.

Figure 2: Depiction of how the sample-based mean speed is

calculated for a given time interval ∆t, where T indicates the

set of all time steps within ∆t. Each of the orange crosses

indicates a sample recorded at the ﬁrst time step.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

100

Usually, one examines separate spatial and temporal

intervals. Since the trafﬁc state is a highly ﬂuctuat-

ing measure precise estimations can only be made on

a local scale, i.e., for certain time intervals (typically

5 min to 30 min) and on given streets or street seg-

ments. These intervals can be referred to as aggrega-

tion intervals and are indicated by the orange boxes.

Of course, if one has access to complete trajecto-

ries, the estimation of the average speed for a spatio-

temporal interval would be as easy as calculating the

arithmetic mean of all respective average speeds as

shown in Equation (1).

V =

∑

i=0

¯v

∑

i=0

∆s

∆t

(1)

In reality, however, one deals with incomplete, po-

tentially erroneous data and has to apply measures

to ﬁll in gaps and correct data. For position mea-

surements, the most prominent ﬁx is to apply Map-

Matching(Brakatsoulas et al., 2005), which will ﬁx

Global Navigation Satellite System (GNSS) positions

to the closest position on a digital map.

2.1 Interpolated Approaches

A conventional approach to mitigate the effect of in-

complete FCD is to apply spline-based interpolation

on the trajectory of the CVs. Yoon et al. (Yoon et al.,

2007) establish the Temporal Mean Speed which is

also visualized in Figure 1. The described method

applies polynomial spline interpolation to reconstruct

vehicle traversals. Based on the resulting splines the

average velocity is calculated for each vehicle accord-

ing to Equation (2).

temporal

= ¯v =

∆x

∆τ

(2)

Even though this method delivers good results when

using a large percentage of CVs and less frequent

transmission of FCD, it quickly suffers in perfor-

mance when decreasing market penetration, due to

unsampled road segments.

2.2 Sample-Based Approaches

Opposing interpolated methods, it is also possible to

treat received samples individually and calculate an

averaged speed estimation that way. One such ap-

proach is displayed in Figure 2 which we label as the

Sample-based Mean Speed. Here, we slice the spatio-

temporal aggregation interval into equally sized time

chunks called T . First, we average all speed values

received for each time chunk T . Afterward, the arith-

metic mean for the entire aggregation interval ∆t is

calculated following Equation (3).

sample

∑

t=1

(

∑

α=1

(t)) (3)

While some information may be lost due to not in-

specting trajectories, this approach offers the ﬂexibil-

ity of integrating samples measured using other vehi-

cles’ perception sensors. Using this approach allows

the integration of information from non-connected ve-

hicles into the TSE, which otherwise would have only

passively affected measurements. Figure 3 exemplar-

ily highlights where the inclusion of perceived vehi-

cles can be useful. These cases can roughly be cate-

gorized as follows:

1. Flowing trafﬁc: Depending on the direction the

perception sensors are facing, it is possible to add

samples from vehicles moving in the same direc-

tion ( ˆ= ﬂowing) as the ego vehicle on all lanes of

the driving direction. This information is less rel-

evant for speed estimates, as other vehicles likely

drive at similar speeds.

2. Oncoming Trafﬁc: Front-facing perception sen-

sors allow the integration of all vehicles on op-

posing lanes if there are no obstructions in the

way. This information is highly valuable because

it gives insights into the trafﬁc state on opposing

lanes correlates only minimally.

3. Intersecting Trafﬁc: The speed information gath-

ered at intersections has to be treated carefully as

it rarely reﬂects the trafﬁc state in its “ﬂowing”

state but more likely vehicles accelerating after

waiting at a red light. Also, intersecting vehicles

can easily be occluded.

Figure 3: Illustration of additional samples acquired

through the use of perception sensors. The black vehicle

is a CV, while the orange vehicles can only be considered in

the TSE due to the additional sensors.

Trafﬁc State Estimation on Urban Roads Using Perception-Enriched Floating Car Data

101

3 SIMULATION APPROACH

As required sensor technologies are not broadly

adapted yet and ﬁeld tests are costly, we conduct our

examinations in a simulated environment. Thus al-

lowing us to implement and emulate all requirements

for a complete system test in a cheap and fast man-

ner. To ensure a degree of accountability and transfer-

ability we base the tests on domain-leading simulators

that have independently stood the test of time.

3.1 Holistic Simulated Vehicle

Perception and Data Fusion

For this paper, we do not aim toward mimicking a spe-

ciﬁc sensor technology and hardware requirements

for our vehicle perception. Instead, we utilize the co-

simulation framework Eclipse MOSAIC and its holis-

tic perception model presented in (Protzmann et al.,

2022). The perception model evaluates vehicle detec-

tions based on a circle sector deﬁned by a radius r and

an angle γ (see Figure 4).

While the applied model is a simpliﬁcation of

the perception task, it enables efﬁcient simulation

even of large-scale urban scenarios like the BeST-

scenario(Schrab et al., 2022) and allows for assessing

the general potential of the proposed TSE approach.

Coordinating data from many error-prone sources

poses a challenge in real-world applications. Noisy

measurements, timing offsets, and network obstruc-

tions (delay, jitter, package loss) hurt the fusion of

samples. We deploy a server-side fusion of data that

fuses samples by using the unique identiﬁers, which

are sent as part of the FCD. Furthermore, we disre-

gard timing and network effects entirely.

- detected - not detected

Figure 4: A simple depiction of the perception model ap-

plied in this paper using a ﬁeld-of-view ﬁlter. Orange vehi-

cles are detected while purple vehicles are not.

Realizing the system with these simpliﬁcations

means we can only make limited prognoses on the

performance of a real-world implementation. How-

ever, since we aim to evaluate the potential of a

perception-enriched FCD, we will disregard these ob-

structions in this initial study. We furthermore, disre-

gard privacy critical considerations but acknowledge

the importance of secure obfuscation of user data.

3.2 Extended MOSAIC Applications

Simulation-based testing requires a strong foundation

of realistic models to yield reliable results. There-

fore, we utilize the well-established simulation frame-

work Eclipse MOSAIC(Schrab et al., 2023) (see Fig-

ure 5), which is loosely based on the IEEE standard

for High-Level Architecture (HLA)(Dahmann et al.,

1997). MOSAIC couples many state-of-the-art FOSS

simulators (Free and Open-Source Software) from

different domains and also comes bundled with its

own. One of the strong suits of MOSAIC is its power-

ful application simulator which enables fast testing of

smart mobility applications, including V2X Commu-

nication via ITS-G5 and LTE/5G, autonomous vehicle

perception, and e-mobility.

The applications utilized for this work are built on

the application simulator and extend an openly avail-

able application suite (github.com/mosaic-addons/

trafﬁc-state-estimation), previously developed and

published. These applications have been built to be

easily extensible and allow for the integration of dif-

ferent sensor technologies on the vehicle side as well

as extensibility in terms of TSE metrics on the server

side.

Interaction

Management

Eclipse MOSAIC Runtime Infrastructure

Time

Management

Federation

Management

Eclipse SUMO

TraCI

Sumo Ambassador

Eclipse MOSAIC Cell

Eclipse MOSAIC

Application

(b)

(a)

BeST-Scenario (Segment)

Figure 5: A rundown on Eclipse MOSAIC’s architecture.

(a) Runtime Infrastructure with the core tasks of federation,

and time and interaction management, enabling a synchro-

nized simulation. (b) Insight into the coupling of trafﬁc sim-

ulator Eclipse SUMO including the BeST-scenario.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

102

A basic schematic of the system design is depicted

in Figure 6. The system is made up of two appli-

cations, one running on all CVs and one running on

a centralized server. On the vehicle side, the FCD

Transmitter Application will periodically collect FCD

samples and transmit them in a collection called FCD

Update to the TSE Server using MOSAIC’s bundled

Cell simulator. In addition to conventional solutions,

we extend the FCD with a list of all surrounding ve-

hicles without any pre-processing for each connected

vehicle (CV) using MOSAIC’s aforementioned per-

ception model. The vehicle application can be conﬁg-

ured with variable sample collection and update trans-

mission times as well as perception ranges and angles.

At the heart of the server lies the TSE Kernel,

which is responsible for receiving the FCD, extract-

ing traversals, and supervising conﬁgured FCD Pro-

cessors. Previously, the Spatio-Temporal Trafﬁc Met-

rics Processor has been established which calculates

the temporal and spatial mean speed as well as the

Relative Trafﬁc Status Metric(Yoon et al., 2007) from

conventional FCD. The results from this processor

are later used as a baseline. In a new development, we

implemented the Perception Metrics Processor which

additionally considers the added data from perception

sensors. This processor aggregates the extended FCD

in a two-stage approach. First, all retrieved samples

are buffered using the time of collection as an index.

FCD-Update

FCD-Record

- position

- speed

- heading

TSE Server

Spatio-Temporal Traffic Metrics Processor

Perception Metrics Processor

- process FCD-Updates

- manage Processors

TSE Kernel

FCD Processors

Map

Database

Traffic

Information

Services

FCD-Update

FCD-Record

- position

- speed

- heading

- perception data

Figure 6: A simpliﬁed system overview of the applications.

The top part depicts the structure of the server application

and the bottom part shows how vehicles record FCD sam-

ples. The major new developments are marked in purple.

After a conﬁgurable delay, samples collected for the

same vehicle at the same time will be averaged (i.e.,

being perceived by multiple other vehicles). We con-

sider all samples as sharing the same truth regardless

of their origin. After the conﬁgured delay, the ag-

gregated samples for each vehicle will be moved to a

secondary buffer that aggregates the samples per net-

work edge, thus generating a mean speed estimate for

each edge and time stamp. The results from this ag-

gregation are then periodically persisted in a database

for further post-processing and evaluation.

In a real-world deployment, the metrics would be

relayed to Trafﬁc Information Services to be accessi-

ble by end users. However, the evaluation of the data

is conducted entirely ofﬂine, after simulation.

4 EVALUATION

In previous work, we presented a calibrated large-

scale urban scenario for the city of Berlin(Schrab

et al., 2022) for the trafﬁc simulator Eclipse

SUMO(Lopez et al., 2018) pre-conﬁgured to be used

in-hand with MOSAIC. The scenario comprises an

800 km

road network, 24 hours of vehicular road

trafﬁc, and 2.25 million individual trips. For the eval-

uation of this paper, we opted to use the bundled sub-

section demand in the Charlottenburg area (see Fig-

ure 7), containing 200,000 trips, reducing simulation

time immensely.

Utilizing MOSAIC’s Mapping Ambassador, we

can equip variable percentages of the BeST trafﬁc

with our applications (Section 3.2). We assume per-

fect perception capabilities and conﬁgured vehicles

with a 100 m viewing range and an angle of 360

◦

, and

also presume perfect communication capabilities ig-

noring effects of delay, packet loss, and jitter.

Figure 7: A map of Charlottenburg indicating the three

street segments under examination.

Trafﬁc State Estimation on Urban Roads Using Perception-Enriched Floating Car Data

103

Table 1: Key markers of the inspected road segments.

street segment length #lanes speed limit signalized

Hofj

agerallee 399.96 m 3 50

Grunewaldstraße 185.37 m 2 50

yes

Bleibtreustraße 182.34 m 1 30

yes

As we are trying to improve on conventional TSE sys-

tems, which rely on 5 % to 10 % market penetration,

we will further investigate rates below this threshold

range. Additionally, we only look at the times from

6 am to 10 pm as the trafﬁc volume outside this time

window is insigniﬁcant, and a TSE becomes obsolete.

In all following plots, we resample results into 15 min

intervals using the arithmetic mean for mean speeds

and summing up collected sample numbers.

4.1 Comparative Examination

In this ﬁrst evaluation, we focus on the street seg-

ments on the Hofj

agerallee, the Grunewaldstraße, and

the Bleibtreustraße, highlighted in Figure 7. These

street segments were also chosen in our previous

work, due to exhibiting different characteristics that

potentially impact the resulting mean speed estimates

(see Table 1). The segment on the Hofj

agerallee

is a long segment with 3 lanes and no trafﬁc light

at the end resulting in a highway-like layout. The

Grunewaldstraße can be classiﬁed as a larger urban

street, with 2 lanes and a signalized end at the in-

spected segment. Lastly, the Bleibtreustraße is a

lowly frequented, almost residential one-lane road

that is also signalized.

Initially, we compare how the novel sample-based

mean speed performs against conventional FCD-

based approaches such as the temporal mean speed.

For this, we create time-series plots for the penetra-

tion rates of 1 %, 2 %, 3 %, 5 %, 10 %, and 100 %

as a baseline. Subsequently, we acquire plots in Fig-

ures 8, 10 and 12. In addition to the sample-based

mean speed (orange) and the temporal mean speed

(purple), we also plot the ground truth (GT) provided

by SUMO as a baseline (gray). Furthermore, we cre-

ate bar charts for the received samples on the server

side in Figures 9, 11 and 13, to compare the received

data based on applied market penetration for each the

extended FCD (left-hand side) and the conventional

FCD (right-hand side).

First, by comparing the temporal mean speed with

the captured ground truth we can revalidate results

from previous research. At 10 % and even 5 % market

penetration, FCD-based TSE still delivers usable re-

sults on the larger Hofj

agerallee and Grunewaldstraße

but fails on the smaller Bleibtreustraße. Interestingly,

on the Hofj

agerallee (Figure 8), the temporal mean

speed still manages to deliver usable results when de-

ploying only 1 % of CVs and only a few intervals go

completely unsampled. This is mainly due to two rea-

sons, (a) the Hofj

agerallee is highly frequented and

thereby likelihood of at least one CV-traversal every

15 min is fairly high, and (b) the variance of the real-

izable mean speed is low throughout the day and be-

tween different vehicles, which means that even one

traversal can capture the mean speed accurately. The

results for the Grunewaldstraße (Figure 10) start to

degrade at higher market penetrations with many un-

sampled intervals at even 3 % and more outliers. We

see increased over- and underestimations of the mean

speed due to red trafﬁc light phases either being cap-

tured or missed when fewer traversals are being ag-

gregated in each interval. Lastly, the temporal mean

speed fails to capture the mean speed at penetration

rates of 5 % and lower for the Bleibtreustraße (Fig-

ure 12), as too few traversals are captured.

Next, looking at the resulting time series for

the sample-based mean speed, some advantages but

also some downfalls can be established. For the

Hofj

agerallee (Figure 8), a clear improvement is ap-

parent compared to the temporal mean speed. Not

only are we able to sample every interval even at 1 %

penetration rate, but we also improve general consis-

tency, capturing virtually zero outliers. The reasoning

for this is similar to the reasons given for the good

performance of the temporal mean speed. Capturing

only a few additional samples of steady-ﬂow trafﬁc

on a large, highly frequented road segment, allows for

an improved TSE because most of the collected sam-

ples can be assumed to reﬂect the mean speed of the

entire segment decently well. This effect cannot be

observed in the same magnitude for the Grunewald-

straße and Bleibtreustraße. On the Grunewaldstraße

(Figure 10), it is apparent that we can capture more

and less ﬂuctuating speed estimates using the sample-

based mean speed even with penetration rates of about

2 %, meaning that the additional data from percep-

tion sensors do improve the overall quality of the es-

timates. Nonetheless, especially at lower penetration

rates, the deviations from the ground truth are larger

compared to the results on the Hofj

agerallee.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

104

Figure 8: A comparison of the aggregated sample-based mean speed against the temporal mean speed on the Hofj

agerallee.

The values are measured for different penetration rates and averaged for 15 min intervals.

Figure 9: Comparison of received data amount depending on the rate of connected vehicles on the Hof

agerallee summed-up

within 15 min intervals. Left: The number of individual samples (i.e., the sum of vehicles per road segment each time step)

used for the sample-based mean speed using conventional FCD plus perception. Right: The number of extracted traversals

used for the temporal mean speed, relying solely on conventional FCD.

Trafﬁc State Estimation on Urban Roads Using Perception-Enriched Floating Car Data

105

Figure 10: A comparison of the aggregated sample-based mean speed against the temporal mean speed on the Grunewald-

straße. The values are measured for different penetration rates and averaged for 15 min intervals.

Figure 11: Comparison of received data amount depending on the rate of connected vehicles on the Grunewaldstraße summed-

up within 15 min intervals. Left: The number of individual samples (i.e., the sum of vehicles per road segment each time step)

used for the sample-based mean speed using conventional FCD plus perception. Right: The number of extracted traversals

used for the temporal mean speed, relying solely on conventional FCD.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

106

Figure 12: A comparison of the aggregated sample-based mean speed against the temporal mean speed on the Bleibtreustraße.

The values are measured for different penetration rates and averaged for 15 min intervals.

Figure 13: Comparison of received data amount depending on the rate of connected vehicles on the Bleibtreustraße summed-

up within 15 min intervals. Left: The number of individual samples (i.e., the sum of vehicles per road segment each time step)

used for the sample-based mean speed using conventional FCD plus perception. Right: The number of extracted traversals

used for the temporal mean speed, relying solely on conventional FCD.

Trafﬁc State Estimation on Urban Roads Using Perception-Enriched Floating Car Data

107

Figure 14: Relation between market penetration and received data points plotted relatively to the GT (i.e., 100 % penetration

rate) indicated for penetration rates of 1 %, 2 %, 3 %, 5 %, 10 %, 50 %, and 100 %. Left: Plot for the extended FCD. Right:

Plot for the conventional FCD. Note, the connecting lines are only added for visualization purposes.

With estimates gathered at 1 % only providing

marginally better results compared to the temporal

mean speed. Looking at the smallest of the three

streets, the Bleibtreustraße (Figure 12), some inter-

esting observations can be made. At 10 % market

penetration, but also at 5 %, 3 %, and 2 %, some out-

liers close to 0 km/h are measured. Going back to

Figure 3 in Section 2, we can assume that an inter-

secting CV perceives one or more vehicles waiting at

a trafﬁc light with low speeds; thus misrepresenting

the estimates for the given time intervals with small

speeds. Multiple measures can be taken to get rid of

these misinterpretations, a simple approach would be

to implement a sample threshold for time intervals,

requiring at least n samples before any aggregation

happens. Still, when disregarding these outliers, we

notice an overall improvement in sample density, de-

creasing the number of unsampled intervals.

After inspecting the resulting time-series dia-

grams, we can shift the focus to the underlying data

by looking at Figures 9, 11 and 13. Preceding any

further evaluation, the difference in signiﬁcance be-

tween Samples for the sample-based mean speed and

Traversals for the temporal mean speed has to be clar-

iﬁed. While a sample is the combined view of a

vehicle’s position and speed at one point in time, a

traversal captures the entire passing of a given road

segment, meaning a traversal encodes much more in-

formation than a single sample. As a consequence,

the magnitude of these values cannot be compared

one-to-one, but rather the percentage of data cap-

tured compared to the ground truth (here meaning the

100 % penetration rate) is important and therefore ad-

ditionally displayed in Figure 14. Without the use

of perception, the traversal amount is directly pro-

portional to the market penetration on all three road

segments, leading to issues on less frequented roads

such as the Bleibtreustraße, especially at smaller pen-

etration rates. Looking at the effect of perception, we

can see that there no longer is a linear relation be-

tween market penetration and the amount of received

data. In fact, on all three roads, the percentages in-

crease over the conventional FCD at all penetration

rates. However, a dependence on the road type is ap-

parent once again; the larger the road, the larger the

number of additional samples received. This is also

very intuitive because the larger a street and its trafﬁc

density are, the more likely it is for additional samples

to be captured by surrounding vehicle perception, sig-

naling increased samples from Flowing Trafﬁc.

In summary, we can see a clear improvement

in estimation quality at lower penetration rates due

to additional samples being captured by the vehicle

perception. Nonetheless, some inconsistencies were

identiﬁed; when too little of a road segment is be-

ing sampled, a misinterpretation of the mean speed

may occur. However, this is more likely to happen

on smaller, less densely traversed roads, for which a

TSE becomes less relevant as these road types rarely

experience congestion or other disruptions. Still, be-

ing able to accurately measure mean speeds on major

road arteries like the Hofj

agerallee with small rates of

CVs (1 % and smaller) is of great value.

4.2 Travel Time Analysis

After having established that the sample-based mean

speed can outperform conventional FCD-based meth-

ods in Section 4.1, especially on high-density streets,

we aimed at an evaluation proxy for the entire road

network. As mentioned previously, mean speed es-

timates can be used as direct input for routing algo-

rithms, but in addition to choosing better routes, more

accurate estimates also enable better calculation of re-

quired travel times. In consequence, if we can pro-

duce overall more accurate travel time estimates com-

pared to the conventional methods, we can globally

assume a better quality of the TSE.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

108

Figure 15: A comparison of the average relative errors between estimated and actual travel times using different mean speed

estimates as input for the routing cost.

To test this, we run the same simulation with the

same demand twice. The ﬁrst iteration was used to

gather speed estimates, following the methods de-

scribed in the prior sections. In addition to the tem-

poral and sample-based mean speed, we included the

ground truth as well as the speed limit for every road

segment. For the second simulation, we developed

simple applications that estimate the travel time for

their given route by summing up the time it would

take to traverse each edge of the route using the de-

scribed speeds. As the realizable speed also has a

temporal dependence, we used the closest available

speed estimates not older than one hour before a ve-

hicle started its tour. If no estimate is available the

speed limit was used as a default. Also note that no

turn costs were considered in the travel time estima-

tion, so a certain bias is to be expected. Results of the

estimations as well as the actual recorded travel times

are written to a ﬁle for later processing, where errors

for all vehicles starting their tours within an interval

are averaged. We calculate the unsigned mean rela-

tive error RME

for each of the travel time estimates

following Equation (4), where t

travel

is the measured

travel time and t

x,est

signiﬁes the applied estimate.

RME

travel

−t

x,est

travel

(4)

The outcome is plotted in Figure 15 at penetration

rates of 100 %, 5 %, 3 %, and 1 % for 15 min intervals.

In these plots, an error of 0.1 signiﬁes that on average

the travel time estimations are wrong by 10 %. We

can see that even when using the measured ground

truth speeds we have an average error of about 15 %,

possibly due to the bias caused by ignoring turn costs.

Furthermore, the largest error can be observed for the

speed limit, which underestimates travel times, es-

pecially with increased trafﬁc density throughout the

midday. For estimates generated using the temporal

and sample-based mean speed, it can be observed that

for all measured penetration rates they both outper-

form the conventional method of using the speed limit

as routing cost. Interestingly, they also outperform

estimates generated using the ground truth; however,

this phenomenon likely occurs due to the inherent bias

introduced by the simpliﬁed cost calculation.

More important is the comparison of the two

estimates with the utilized equipment rate of CVs.

While for both estimations the average error increases

at lower penetration rates, the perception-induced

sample-based mean speed can better compensate for

missing samples and provides better travel time esti-

mates at lower penetration rates. This becomes es-

pecially apparent when looking at the 1 % case where

the temporal mean speed estimates more often default

to using the speed limit thereby increasing the overall

error, while the sample-based mean speed estimates

still provide consistent results. Following these re-

sults an examination of equipment rates of less than

1 % appears to be of interest, unfortunately, these ex-

periments could not be conducted within the scope of

this paper.

Trafﬁc State Estimation on Urban Roads Using Perception-Enriched Floating Car Data

109

It is important to not lift these results out of pro-

portion, particularly because we used simpliﬁed and

over-realistic models for communication and percep-

tion. Nonetheless, the potential upsides of enrich-

ing FCD with data from perception sensors have been

proven and should further be examined.

5 CONCLUSION & OUTLOOK

In this paper, we investigated the research question

of whether a good level of Trafﬁc State Estimation

(TSE) based on Floating Car Data (FCD) can be real-

ized with sufﬁcient precision, even at low equipment

rates, yet with high data quality from perception sen-

sors.

In previous work(Schweppenh

auser et al., 2023),

we offered an introduction to established speed met-

rics for TSE and respective sensor modalities as well

as their upsides and shortcomings. The main aspects

of this work have been reiterated in the ﬁrst part of

this paper. Floating Car Data approaches have been

identiﬁed to be the most common source for current-

day navigation applications. The Temporal Mean

Speed(Yoon et al., 2007) was found to be one of the

core measures to be estimated from FCD traces us-

ing curve-ﬁtted traversals. In contrast to conventional

methods, we inspected how perception sensors could

be used to improve the data quality for the sake of

mean speed estimation. We differentiated three trafﬁc

scenarios, Flowing, Oncoming, and Intersecting Traf-

ﬁc that show different levels of potential gain. Based

on this, we utilized Eclipse MOSAIC’s(Schrab et al.,

2023) perception module and showed how data gath-

ered from surrounding vehicles can be used as addi-

tional input for mean speed estimation in the form of

the Sample-Based Mean Speed.

To evaluate the established measures, we set up

a simulation-based test, leveraging the strengths of

MOSAIC. Building on top of previously published

applications

, we developed a new processor for cal-

culating the sample-based mean speed. By coupling

state-of-the-art simulators and utilizing the calibrated

large-scale BeST scenario(Schrab et al., 2022), we

can generate sensible and transferable results.

Focusing on the area of Charlottenburg, we

ﬁrst conducted a study comparing the temporal and

sample-based mean speed with the ground truth gen-

erated by the trafﬁc simulator Eclipse SUMO. We

furthermore identiﬁed that estimation quality strongly

depends on the road type for both the conventional

https://github.com/mosaic-addons/trafﬁc-state-

estimation

and the perception-enriched FCD as well as the mar-

ket penetration of connected vehicles. Consequently,

we looked at time-series plots for the speed estima-

tions and the amount of data received for three ex-

emplary road segments in dependence on penetration

rates. We were able to validate the proposed ap-

proach of the Sample-Based Mean Speed as a sensible

method to estimate the mean speed. This analysis fur-

ther showed that regardless of the targeted road type

the perception increases the amount of received data,

and thereby improves estimation quality. As a conse-

quence, this indicates that penetration rates of 1 % and

lower can be sufﬁcient to provide reliable mean speed

estimates for larger road segments. Nonetheless, we

also identiﬁed shortcomings and less promising re-

sults for smaller road segments, which we intend to

further investigate in future work.

Evaluating the estimation quality on a global scale

is a non-trivial task. Therefore, we deﬁned the result-

ing travel time estimations as a proxy for the overall

TSE quality. Our results indicated that using percep-

tion data from surrounding vehicles improves average

travel time estimates over conventional FCD, partic-

ularly at lower penetration rates, thereby indicating a

globally improved estimation quality.

Concerning a real-world deployment of the pro-

posed system, some challenges arise. For one, we

heavily simpliﬁed the vehicle perception and sen-

sor fusion. Real sensors suffer from occlusion and

noise, which reduce sample size and quality. Sec-

ondly, we disregarded the constraints of the cellular

network. Cell coverage, timing issues, losses, and

package sizes are not realistically modeled. Nonethe-

less, this study still holds meaning as the listed limita-

tions can either be overcome by improved technology

or only have a minor impact on ﬁnal results.

In conclusion, our results indicate that data from

vehicular perception sensors can be a valuable addi-

tion to the input of Intelligent Transportation Systems

such as TSE and should be further researched. In the

future, we aim to make our results more quantiﬁable

by improving the realism of the applied perception

model, including occlusion and noise. We further aim

to investigate how the deployed estimation techniques

react to more disruptive trafﬁc patterns like conges-

tion, construction sides, and second-row parking.

ACKNOWLEDGEMENTS

This work was supported by the KIS’M project

through the German Federal Ministry for Economic

Affairs and Climate Action under grant 45AVF3001E.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

110

REFERENCES

Blokpoel, R. J., Krajzewicz, D., and Nippold, R. (2010).

Unambiguous metrics for evaluation of trafﬁc net-

works. In 13th International IEEE Conference on

Intelligent Transportation Systems, pages 1277–1282.

IEEE.

Brakatsoulas, S., Pfoser, D., Salas, R., and Wenk, C. (2005).

On Map-Matching Vehicle Tracking Data. In Pro-

ceedings of the 31st International Conference on Very

large Data Bases, pages 853–864.

Dahmann, J. S., Fujimoto, R. M., and Weatherly, R. M.

(1997). The Department of Defense High Level Ar-

chitecture. In Proceedings of the 29th Conference on

Winter Simulation, pages 142–149.

Delooz, Q., Riebl, R., Festag, A., and Vinel, A. (2020). De-

sign and Performance of Congestion-Aware Collec-

tive Perception. In 2020 IEEE Vehicular Networking

Conference (VNC), pages 1–8. IEEE.

Ferman, M. A., Blumenfeld, D. E., and Dai, X. (2005). An

Analytical Evaluation of a Real-Time Trafﬁc Informa-

tion System Using Probe Vehicles. In Journal of In-

telligent Transportation Systems, volume 9, pages 23–

34. Taylor & Francis.

Lopez, P. A., Behrisch, M., Bieker-Walz, L., Erdmann, J.,

otter

od, Y.-P., Hilbrich, R., L

ucken, L., Rummel,

J., Wagner, P., and Wießner, E. (2018). Microscopic

Trafﬁc Simulation using SUMO. In 2018 21st In-

ternational Conference on Intelligent Transportation

Systems (ITSC), pages 2575–2582. IEEE.

Messelodi, S., Modena, C. M., Zanin, M., De Natale, F. G.,

Granelli, F., Betterle, E., and Guarise, A. (2009). In-

telligent extended ﬂoating car data collection. Expert

Systems with Applications, 36(3, Part 1):4213–4227.

Protzmann, R., Schrab, K., Schweppenh

auser, M., and

Radusch, I. (2022). Implementation of a Perception

Module for Smart Mobility Applications in Eclipse

MOSAIC. In SUMO Conference Proceedings, vol-

ume 3, pages 199–214.

Ruppe, S., Junghans, M., Haberjahn, M., and Troppenz, C.

(2012). Augmenting the Floating Car Data Approach

by Dynamic Indirect Trafﬁc Detection. Procedia - So-

cial and Behavioral Sciences, 48:1525–1534. Trans-

port Research Arena 2012.

Schrab, K., Neubauer, M., Protzmann, R., Radusch, I.,

Manganiaris, S., Lytrivis, P., and Amditis, A. J.

(2023). Modeling an ITS Management Solution for

Mixed Highway Trafﬁc With Eclipse MOSAIC. IEEE

Transactions on Intelligent Transportation Systems,

24(6):6575–6585.

Schrab, K., Protzmann, R., and Radusch, I. (2022). A

Large-Scale Trafﬁc Scenario of Berlin for Evaluating

Smart Mobility Applications. In Conference on Sus-

tainable Urban Mobility, pages 276–287. Springer.

Schweppenh

auser, M., Schrab, K., Protzmann, R., and

Radusch, I. (2023). Spatio-Temporal Speed Metrics

for Trafﬁc State Estimation on Complex Urban Roads.

In International Conference on Simulation Tools and

Techniques. Springer.

Wardrop, J. G. and Charlesworth, G. (1954). A Method of

Estimating Speed and Flow of Trafﬁc from a Moving

Vehicle. Proceedings of the Institution of Civil Engi-

neers, 3(1):158–171.

Yoon, J., Noble, B., and Liu, M. (2007). Surface Street

Trafﬁc Estimation. In Proceedings of the 5th Inter-

national Conference on Mobile systems, applications

and services, pages 220–232.

Trafﬁc State Estimation on Urban Roads Using Perception-Enriched Floating Car Data

111