Simulation Environment for Traffic Control Systems Targeting

Mixed Autonomy Traffic Scenarios

Christopher Link

, Kevin Malena

, Sandra Gausemeier

and Ansgar Trächtler

1,2

Heinz Nixdorf Institute, University of Paderborn, Fürstenallee 11, Paderborn, Germany

Fraunhofer Institute for Mechatronic Systems Design IEM, Zunkunftsmeile 1, Paderborn, Germany

Keywords: Traffic Simulation, Traffic Control, Car2X, Mixed Autonomy, Autonomous Vehicles, SUMO, Sensor

Simulation, Traffic Demand Generation, Routing, Traffic Lights, Graph Analysis, Traffic Observer.

Abstract: The development of autonomous vehicles and their introduction in urban traffic offer many opportunities for

traffic improvements. In this paper, an approach for a future traffic control system for mixed autonomy traffic

environments is presented. Furthermore, a simulation framework based on the city of Paderborn is introduced

to enable the development and examination of such a system. This encompasses multiple elements including

the road network itself, traffic lights, sensors as well as methods to analyse the topology of the network.

Furthermore, a procedure for traffic demand generation and routing is presented based on statistical data of

the city and traffic data obtained by measurements. The resulting model can receive and apply the generated

control inputs and in turn generates simulated sensor data for the control system based on the current system

state.

1 INTRODUCTION

Traffic control is a vital part of road mobility,

especially in urban areas. It is required for an efficient

use of the given road network and often has a direct

impact on the traffic situation. Current means of

traffic control e. g., traffic light systems (TLS) or

dynamic speed limitations on selected roads are key

to improve the traffic flow and are therefore currently

subject to optimisation. However, a major drawback

of those systems is their cost and the need for

additional infrastructure, like traffic detectors, to

acquire an accurate picture of the traffic state. TLS,

for example, can achieve significantly better

performance if the traffic situation in their vicinity

can be observed and used to determine the most

suitable control input to handle the current situation

(Malena et al., 2022).

The current developments in autonomous and

connected vehicles offer a great potential to remedy

these limitations and to integrate these vehicles as

agents in traffic control systems themselves. Modern

vehicles monitor their own position and speed

continuously and could share this information using

Car2X technology. Moreover, autonomous vehicles

are dependent on the constant observation of their

environment in real-time to be able to drive without

assistance of a person. For a central traffic control

system this data can give valuable insights in the local

traffic situation close to the respective vehicles.

Additionally, the current road infrastructure already

integrates detectors, like induction loops or radar

detectors, which can provide traffic data on stationary

locations of varying quality. The incorporation of all

these local data sources into a central control system

can be used to obtain a comprehensive and up-to-date

picture of the traffic network’s state, even if only a

small share of the road users is able or willing to

participate in data-sharing.

Our goal, based on these considerations, is to

develop such a traffic control system for future traffic

scenarios and to utilizes the capabilities these systems

would enable. A real-time picture of the traffic state

can further improve TLS performance and enable

route optimisation or rerouting suggestions for

vehicles on the road. Furthermore, cooperative

autonomous vehicles could be used to adapt the

traffic flow speed in order to reduce congestions or

stop-and-go traffic on road sections ahead. The

framework development requires a suitable

simulation environment since the traffic composition,

as described above, and the means for data-sharing

currently do not exist to the required extent.

Link, C., Malena, K., Gausemeier, S. and Trächtler, A.

Simulation Environment for Trafﬁc Control Systems Targeting Mixed Autonomy Trafﬁc Scenarios.

DOI: 10.5220/0011987600003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 367-377

ISBN: 978-989-758-652-1; ISSN: 2184-495X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

367

In this paper, we present a simulation environment

and its components that is the foundation of such a

system. In section 2, a literature overview is given to

set a baseline for the presented research. An overview

over the system and its structure is presented in

section 3. Section 4 comprises the modelling of the

road infrastructure as well as the traffic demand.

Finally, in section 5 the conclusions are drawn, and

the next steps are formulated.

2 LITERATURE OVERVIEW

Modelling a traffic environment and the associated

control system are comprehensive tasks that include

several aspects. In this section, a selective overview

is presented due to the limited scope of this

publication. There are many approaches and tools

resulting from previous research. In (Lopez et al.,

2018), the microscopic traffic simulation tool SUMO

and its framework is presented which is used in this

research. By employing the open-source software it is

possible to model and simulate traffic scenarios in a

realistic way. However, to reproduce the behaviour of

a real traffic system, the mobility demand must be

approximated realistically. An overview over

methods for activity-based demand generation is

given in (Schweizer et al., 2018). Depending on the

available data, the desired output and the scope of the

simulation, different methods can be pursued.

Usually, the approaches derive the traffic demand

from data about the population and its behaviour in

the regarded area. Other structural information, e. g.,

the location and size of schools can also impact the

simulated traffic situation, as investigated in (Ma et

al., 2020). In that research, SUMO was used in

combination with the tool Activitygen to simulate a

realistic traffic environment and the results were

compared to real traffic data. An alternative approach

is used in (Maiorov et al., 2019) by splitting a large

traffic region in multiple sections. Applying a

gravitational model and incorporating structural data

of the region, origin-destination (OD) matrices

describing the traffic flows between the sections are

created and used for route generation. In (Lobo et al.,

2020), a method is presented to create a traffic model

of an urban area using SUMO, Activitygen and an

iterative routing approach. This is based on real traffic

measurements and includes realistic programs for

some TLS.

To obtain the road network, data from

OpenStreetMap (OpenStreetMap contributors, 2022)

can be imported and converted to the SUMO

standards. A topological analysis of such a road

network based on graph theory was performed by

(Henning et al., 2017). Using several metrices, e. g.,

Betweenness Centrality and Closeness Centrality, a

topology-based method was applied to identify

important roads of the network. The results of such an

analysis can be used by control systems or traffic

planers to improve the traffic situation. A similar

graph-based approach was taken by (Ahmadzai et al.,

2019) to rate a city’s road network.

The components of the planned control system are

also based on prior research analysing different

aspects. In (Farrag et al., 2020), information about the

simulated traffic obtained via Car2X technology is

used to identify and react to local traffic incidents

(e. g., blocked lanes). Using this information

subsequent vehicles can reduce their velocities which

leads to a reduction in time loss. This demonstrates

both, the potential of sharing traffic information and

the capabilities of velocity control. This assessment is

supported by the results of (Guo et al., 2020) who

consider a mixed autonomy traffic situation. They

show that through speed harmonization on roads

leading to known bottlenecks, a better traffic flow can

be achieved, provided there is sufficient sensor

coverage in these areas. In (Malena et al., 2021a) and

(Malena et al., 2021b), we present a validated method

to obtain the traffic state of a real-world traffic area in

real-time using stationary detectors. Using this

approach, we were able to control multiple TLS in the

regarded area and to integrate a more suitable target

phase selection for the current traffic situation

(Malena et al., 2022). However, this approach was

limited to a city district of Paderborn and required the

integration of additional sensor systems.

Finally, the subject of vehicle routing is important

for the planned system since it is a key part of the

desired real-time control system, and it is required for

the initial route allocation as well. In (Lazar et al.,

2021), a deep reinforcement learning algorithm is

used for cooperative routing of the autonomous

vehicles in a mixed autonomy environment while

human-driven vehicles rely on selfish route choice.

The research shows that a cooperative approach can

lead to a reduction of travel times even if it is only

applied to a fraction of the vehicles. Similar results

were achieved by (Krichene et al., 2016) indicating

that even a small share of controllable vehicles can be

used to achieve significant improvements.

Furthermore, other possible solutions to remedy the

inefficiency of selfish, non-cooperative routing are

summarised, like pricing congestion or the allocation

of road capacities.

The traffic control system and the simulation

environment presented below are based on the

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

368

research presented here, combining and extending it.

The SUMO-based simulation is complemented by

new models of sensor systems and TLS. Also, a

routing system was developed including results from

topological analysis and real traffic data.

3 SYSTEM OVERVIEW

In order to describe the simulation model, it is

necessary to understand how it is embedded in the

control loop and to formulate the requirements it has

to meet.

Figure 1: System overview.

The structure of the system which is being

developed is depicted in figure 1 and consists of three

major components: A simulation model, a traffic

observer, and a control system. The latter two serve

as the information processing unit which receives

simulated sensor data from the model and in turn

provides appropriate control inputs. In contrast to

that, the simulation model is a substitute for a real

traffic environment and is used to test the control

system. Therefore, it has to encompass all relevant

components which have a meaningful influence on

the system’s behaviour. It must be able to receive and

apply the given control inputs and generate the

required sensor data based on the system’s state. The

controller provides three means to interact with the

traffic system:

A routing system utilizes the knowledge of the

system state to dynamically find optimised paths to

the given destinations. Such a system must weigh

between multiple criteria, such as route length,

expected travel time, traffic density etc. and suggest a

route-change for compatible vehicles. It is assumed

that autonomous vehicles follow these suggestions

while the drivers of other Car2X-equipped vehicles

can reject them which can be modelled using a

probability-based approach. Other vehicles cannot be

controlled by this system directly, however by easing

traffic demand on critical road sections they are also

expected to experience an indirect positive effect.

The second mean of control is also aimed at

Car2X-equipped vehicles and utilizes the ability to

send them respective target velocities for the road

sections they are located on. If the estimated target

velocity is below the current traffic flow speed it can

be enforced by autonomous vehicles and cooperating

human-driven vehicles. Using consensus algorithms

for a purely autonomous traffic showed that it is

possible to achieve a more homogenous traffic flow

and to reduce undesirable effects like stop-and-go

traffic if the target velocities are chosen appropriately

(Mertin et al., 2020).

Since TLS have a great influence on the traffic

flows, the final control structure is a system to

optimise their performance. Based on the estimated

traffic state on the roads in the vicinity of each TLS,

the waiting times of the affected vehicles and their

vehicle types (if known), a target phase and the

desired switching time is to be calculated and applied

to the traffic system. In our prior research, we have

developed an approach based on Model Predictive

Control (MPC), which is able to improve the

performance of TLS significantly compared to

control systems currently in use (Malena et al., 2022).

An integration of this control approach is therefore

planned for this system as well.

All those presented control systems prerequisite a

comprehensive and up-to-date knowledge of the

current system state in order to provide effective

control inputs. To achieve this, an observer is

currently under development, which processes and

utilizes the data obtained by various sources in the

simulation model. A probability-based traffic

simulation is used to extrapolate the estimated system

state and is continuously updated and corrected by the

incoming sensor data. The accuracy of the estimated

system state of certain areas of the road network

therefore depends on the availability of recent sensor

data. Further details about the observer are subject for

a future publication as soon as implementation and

further tests are completed.

Based on these interactions with other system

components, it is possible to formulate several

requirements for the simulation model: The road

Simulation Model

Control System Traffic Observer

Traffic Dynamics

(SUMO)

Traffic

Demand

Road

Network

AV / Car2X-Vehicles

TLS-

Model

Sensors

TLS-Phases

True State

Sensor Signals

Interaction

Traffic Light Signals

Routing System

Velocity Control

TLS Controller

Sensor Fusion

Probability-

based Traffic

Simulation

Velocity

Targets

Routes

Observed

State

Car2X Data

Simulation Environment for Trafﬁc Control Systems Targeting Mixed Autonomy Trafﬁc Scenarios

369

network has to be selected and provided to the various

system components. The vehicle dynamics must be

modelled containing relevant functions like car-

following-models or lane-change-models. Also, the

behaviour of the different vehicle types has to be

specified, especially the behaviour of autonomous

vehicles. This includes means to set the desired target

velocities and routes. TLS must be integrated in the

road network and the respective controllers must

ensure that the given target phases are applied in a

realistic way. Additionally, the sensors have to be

modelled to generate the required data for the traffic

observer. To test the system under realistic conditions

the traffic demand must be determined and used to

generate appropriate trips which in turn are the basis

to calculate realistic initial routes for the vehicles in

the simulation. In the following section, the

simulation model is described in detail.

4 MODELLING PROCESS

In this section, the simulation environment and its

relevant components are addressed and discussed,

beginning with the selection of the road network and

the model basics. Subsequently, the modelling of the

traffic infrastructure is presented and followed by the

method used for traffic demand generation. The

simulation environment is built on the traffic

simulation software SUMO which is an open-source

tool maintained mainly by the German Aerospace

Center. It is based on a microscopic traffic model and

includes several sub-models e. g., for lane-changing,

car-following behaviour or the reaction of the drivers

to TLS. Therefore, it provides a suitable and

extendable base to meet the requirements listed

above. The system components rely on multiple data

Figure 2: Data processing for the simulation model.

sets to perform their respective tasks. An overview

over the processes needed to generate these data sets

is depicted in figure 2 and will be referred to in the

following sections.

4.1 Traffic Infrastructure

4.1.1 Road Network

The simulation model is based on a real-world traffic

environment to demonstrate the applicability of the

control system to existing road networks. The city of

Paderborn, Germany, and parts of the surrounding

area (see figure 3) serve as a template for the model.

It includes over 960 km of roads (counting both

directions separately) and consists of a wide variety

of road types, from an Autobahn (highway) to

residential streets. Also, there are 137 TLS in the

network which are also considered for the simulation.

Some roads are exclusively for public transportation

and authorities.

The foundation of the road network was imported

from OpenStreetMap (OSM) using the import tool

Netconvert provided by the SUMO toolkit. Although

it provides a useful basis for the model it is far from

being directly deployable for simulation and requires

extensive manual corrections to compensate for

incorrect data. This especially applies to junctions

and the correct definition of the lane connections

indicating which target lanes can be reached from

which origin lanes.

Figure 3: Extract of the selected road network of the city of

Paderborn over the corresponding OSM-map layer

(OpenStreetMap contributors, 2022).

Java Road

Network

Topology

Analysis

SUMO Road

Network

Road Priority

Analysis

Road Priority

Assessment

Traffic

Demand

TLS Config

Data

Initial Vehicle

Routing

SUMO-

ActivityGen

Trips

(Origin-

Destination)

Structural

Data of

Paderborn

Measured

Traffic

Data

Derived

Data

Process

OpenStreet-

Map

SUMO-

Netconvert

Base

Data

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

370

In order to provide topological network

information to the components of the system (i. e., the

observer and the controller), a tool was developed to

analyse the resulting SUMO road network file and to

create a corresponding graph-representation. This

includes the identification of areas of influence (e. g.,

for TLS based on the distance to them) as well as the

identification of parallel and counter-directional

roads since these are relevant for the control system

design. Another field of application for topological

network information is to integrate the TLS, which is

described below.

4.1.2 Road Priority Assessment

The road network encompassed by the simulation

model has an extensive expanse, however many roads

lead through residential areas, the town centre,

industrial areas, or other regions which are not

intended to be a central vein of transportation.

Especially for the routing-based control system it is

important to consider this since extensive routing

through these areas might cause problems or

discontent among the residents. The importance of a

road is represented by OSM in a priority class which

is based on its role (e. g., highway, federal road), the

number of lanes, etc.

On the other hand, it is important to identify

which roads are most valuable for the traffic system

due to the location or connections to other roads. A

mean to incorporate this is to introduce a numerical

priority rating for each road section as a combination

of several criteria. An established method to rate the

importance of a node in a graph is the Betweenness-

Centrality. It is an indicator of how often a given node

is part of the least costly connection between any two

nodes in the graphs. To apply this to the road network,

it is converted to a line graph that allocates a node to

each road. The connecting edges are created based on

the reachable follow-up roads at each junction. The

cost associated with each edge is set using two

different attributes of the roads. This leads to two

independent graphs: One weighed with the length of

the respective roads and the other with the free travel

time on them (i. e., the length divided by the speed

limit). These attributes were selected because both

play an important role for drivers’ route decisions and

should therefore both be regarded. Consequently, the

Betweenness Centrality is calculated for both cases.

To combine the priority rating 𝑟

,

for road 𝜏

from OSM with the values obtained by the graph

analysis using Betweenness Centralities based on

distance and travel time (𝑟

,

,𝑟

,

), an optimisation

problem is formulated. The goal is to find the

weighting 𝑤∈ℝ



for the different rating approaches

that minimizes the quadratic deviation between a

linear combination of the ratings and actual traffic

data 𝑣



for 𝑛 measurement locations available in

Paderborn with

𝑤=argmin

∈

ℝ



∑

𝑣



−

𝑟

,

𝑟

,

𝑟

,





⋅𝑤







(1)

In Paderborn, these measurements were acquired by

a traffic count from 2018 and include the vehicles

detected per day at about 300 locations within the

network. The resulting weighting 𝑤 is used to

calculate a single value in the range 0 to 1.

4.1.3 Traffic Light Systems

As stated above, TLS are a key part of urban traffic

environments since they exert an immediate influence

on the traffic flows. SUMO supports the integration

of TLS and the simulated traffic participants abide to

the signal lights. However, it is up to the user to

ensure that the TLS behave as expected. This is not

limited to the selection of the TLS-phases (which is

the responsibility of the control system) but also to

the way they are implemented, e. g., regard the

minimum green durations and transition times and

avoid incompatible signal combinations. Depending

on the junction’s geometry and the desired phase

transitions these details can have a significant impact

on the transition times and therefore should not be

neglected. To ensure that TLS in the simulation

comply with, the official guidelines for TLS in

Germany, RiLSA (RiLSA, 2015), a controller was

developed that commands the signal lights based on

the current TLS state and the desired target phase. It

ensures the adherence to the yellow signal change

time, the minimum green time and red clearance for

conflicting signals. The basics of the underlying

controller logic applied to each signal of a TLS are

depicted in figure 4.

Each TLS-signal may control multiple lanes, but

a lane is not limited to a single signal. E. g., a right-

turning lane can be released by the ‘main’ green light

or by a right-arrow green signal provided it has no

dedicated red light. Depending on the current state of

each signal and the selected target phase, a series of

checks is performed to decide whether a state

transition is permitted or must be postponed. This

might even result in an active signal to be turned to

red temporarily in order to allow a dependent signal

to switch to green. The mutual relationships and

dependencies of the signals depend on the geometry

and conflict areas of the intersection. Lanes with no

Simulation Environment for Trafﬁc Control Systems Targeting Mixed Autonomy Trafﬁc Scenarios

371

conflicting points and active signals can be controlled

independently. They impose no restrictions for the

controlling signals while lanes from a crossing street

must not receive green if there are any conflicting

areas.

Figure 4: TLS controller logic (Malena et al., 2022).

An automated analyses is performed for each

junction comprising TLS which generates plausible

datasets to setup the simulation and the controller

based on geometrical features. Depending on the

incoming and outgoing lanes and the connections

between them, a TLS setup is selected, and matching

signals and phases are generated. Also, restrictions

like minimum green times and transition times are set

based on that. For TLS with unusual setups or to

incorporate actual phase plans the configuration

datasets can be adjusted manually to ensure a realistic

behaviour.

4.1.4 Traffic Sensors

In a real-world scenario, the sensors are a vital part of

the traffic control system since they are the only

source of information about the current traffic state.

To test the developed system under realistic

conditions, it is therefore required that suitable data

packages are generated by ‘virtual’ sensors based on

the known system state of the simulation model. In

this research, there are five different sources of

information modelled. They vary in terms of the

provided data as well as the time at which the data is

shared. The sensor types with the corresponding

information they provide are listed in table 1.

Table 1: Simulated sensors and their provided information.

Data Source Transmitted Information

Car2X-equipped

vehicles

- Own position

- Own velocity

- Own route (if available)

- Own vehicle type

Autonomous

vehicles

- All Car2X-vehicles’ data above

- Nearby vehicles’ positions

- Nearby vehicles’ velocities

- Nearby vehicles’ types

Induction Loops - Time of detections

- Current occupation status

Radar Detectors - Time of detections

- Detected vehicles’ velocities

- Detected vehicles’ types

Aggregated data

sources

- Average traffic density over a

given time span (delayed)

Induction loops are placed at the stopping lines of

each incoming lane at a TLS. Furthermore, for TLS

containing multiple lanes per direction, additional

induction loops are set up to 40 meters ahead of the

junction, as this is a common setup in Paderborn.

Induction loops are prone to errors especially if

crossed by small vehicles which do not inflict a huge

impact on the inductivity of the sensor. To model this,

a vehicle type-dependant probability is defined to

determine if the crossing of a road user is actually

registered by the sensor. Radar detectors are also

stationary and placed manually on the road network.

They are intended to augment the data collection

efforts on road sections which do not feature

induction loops but exhibit a sufficiently high traffic

volume that would justify an installation of such a

device. Radar detectors provide more reliable

measurements than induction loops. Also, they are

able to gather additional information like the crossing

vehicles’ velocities and vehicle types (which is also

affected by a type-dependent misclassification

probability). To generate the sensor data packages,

the set 𝐿

,

containing all vehicles 𝑖 on lane 𝑧 at the

current time step 𝑘 is considered. For each vehicle

being on the lane for one of consecutive time steps,

i. e., 𝑖∈𝐿

,

∪𝐿

,

, the following cases are

checked using the vehicles’ positions 𝑥

,

on their

respective lane and the positions 𝑝



of the lanes’

respective sensors:



𝑖∈𝐿

,



∧



𝑥

,

≥𝑝





∧



𝑖∉𝐿

,





𝑖∈𝐿

,



∧



𝑥

,

≥𝑝





∧



𝑥

,

<𝑝





(2)



𝑖∈𝐿

,



∧



𝑥

,

<𝑝





∧



𝑖∉𝐿

,



If any of the three conditions apply, a sensor crossing

was determined. Consequently, the corresponding

data is read from the known true simulation state and

Min.

green time

reached?

*Conditions:

 No conflicting phase active?

 Red clearance time for all conflicting phases reached?

 Are above conditions also met for all dependent signals?*

Other

green signal

blocked?**

All conditions

met?*

Initiate

switch to r

Wait for next time step

Initiate

switch to g

State: g/u

Target: r

State: g/u

Target: g

State: r/y

Target: g

false

Get current

and target

signal state

**To prevent a permissive signal from becoming green if it would receive

the right-of-way over another green signal (prohibited by the RiLSA)

Signal states

g:Green r:Red

y: Yellow u: Red & Yellow

false

true

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

372

modified based on the misdetection and

misclassification probabilities according to the type

of sensor. The data package is provided to the traffic

control system without additional delay since the

transmission time is neglectable compared to the

simulation step size (Δ𝑡=1 𝑠) when using a suitable

transmission protocol.

Due to the usage of Car2X communication

technology by some traffic participants, these can be

utilized as moving data sources. It is assumed that

they share their current position, velocity, and the

route they are on, if set. Additionally, autonomous

vehicles provide information which they gather from

their environment. Thus, the position, velocity and

type of nearby vehicles are transmitted as well.

Especially at junctions this can concern vehicles on

multiple lanes, therefore a comparison of the state of

the current lane is not sufficient. To preselect the

vehicles which might be in range of autonomous cars,

the road network is divided into a grid with several

100m-by-100m fields. Each vehicle is allocated to the

corresponding field using its coordinates once per

time step. Thus, to determine the nearby vehicles it is

sufficient to check the distance to the vehicles in the

current grid-field and all neighbouring fields. This

reduces the number of checks to be performed

significantly. Similar to the stationary sensors it is

assumed that the data packages from Car2X-equipped

vehicles are gathered each time step with no

additional delay.

The final method to get information about the

traffic are sources that aggregate data using different

sensors before sharing it. These do not provide data

of individual vehicles but instead estimate the traffic

density on road segments of the network. Data like

this can be gathered by using cell phone information

e. g., via navigation apps or detecting and counting

nearby Bluetooth devices. The aggregated data is not

available in real-time but can still be important for a

traffic control system, especially for routing purposes

and areas of the network without a great sensor

coverage. While the collection methods are not

modelled in detail, the aggregated traffic density over

a certain time period is modelled using a moving

average of the vehicle count for each lane and

delaying it further.

4.2 Traffic Demand

The fixed road infrastructure and the sensor systems

are the base of the simulation model. However, to

perform simulations the traffic itself has to be

modelled as well. This includes the definition of the

vehicle types and their respective behaviour as well

as the generation of the traffic demand to fill the road

network in a plausible way.

In the SUMO traffic simulation eight different

vehicle types are used, e. g. passenger-cars or busses.

Additionally, there are autonomous variants of most

of the types. For human-driven vehicle types the

parameters are configured to randomly deviate from

an ‘ideal’ driving behaviour. This may result in

divergences from the speed limits, not keeping a

sufficient distance to the preceding vehicle, or an

impatient behaviour at junctions regarding the right-

of-way. For autonomous vehicles these deviations are

disabled since a computer-driven car would not

deliberately violate the traffic rules and cannot be

distracted. Moreover, autonomous vehicles can be

controlled by the traffic control system to some

degree.

4.2.1 Trip Generation

The goal of the traffic generation is not to recreate the

exact traffic which is present in Paderborn on a given

day (this would require to monitor each traffic

participant individually), but to create a plausible

traffic situation that resembles the real traffic. The

first step is to generate the trips that are to be carried

out during the simulation. A trip defines the origin

and the destination a vehicles’ route has to connect as

well as the departure time. Therefore, the entirety of

all trips represents the demand for mobility in the

network without prescribing how it is realized. The

(initial) routes the vehicles should take are

determined in a subsequent step (see section 4.2.3).

SUMO already includes the application Activitygen to

generate trips based on a given road network and

additional information about the environment which

must be provided externally. These include statistical

data about the population in Paderborn, like

population count, demographics, employment rate,

car ownership rate, etc. Also, information about the

number of incoming and outgoing commuters as well

as an approximate distribution of the usual working

hours are given. This information was mainly

obtained by publicly available data published by the

City of Paderborn.

Furthermore, information directly related to the

road network was provided that includes the position

and size of schools in the regarded area, since they are

a common destination for many trips at certain times

of the day. Also, the main roads leading in and out of

the road network were specified including the number

of vehicles traveling on them each day. For most of

the relevant roads this information could be obtained

from the traffic count mentioned above. The traffic

Simulation Environment for Trafﬁc Control Systems Targeting Mixed Autonomy Trafﬁc Scenarios

373

which is generated from and to a certain road is

dependent on the number of residents that live there,

and the number of workplaces located nearby. This is

not uniform for all roads and can vary significantly

depending on the location. As there are more than

40,000 road sections in the road network, a manual

setting is not practical, and an alternative method was

used.

To more efficiently allocate numerical values to

the different road edges an image-based method was

developed. It uses the RGB colour channels of a new

layer which is added to the map displayed in figure 3

to encode the required information. Consequently,

areas on the map can be marked according to their

position and function with different colour intensities.

A transformation function is used to map each

network coordinate to the corresponding pixel of the

image. Thus, for each road edge a central coordinate

is determined, the corresponding pixel is selected, and

the respective value is read based on the red

component of its RGB value using a conversion

factor. Note that this is a relative value and the total

number of workers and inhabitants to be allocated is

calculated based on the statistical data mentioned

above. Since there is no data available for Paderborn

that shows the population density or the density of

workplaces in the level of geographical detail

required, the areas on the map are marked based on

their urban development and function. Residential

areas receive a higher-than-average rating for the

number of inhabitants while exhibiting only a limited

number of workplaces. For industrial areas on the

other hand an inverse structure is defined, and the city

centre exhibits high values in both regards. Using this

input data Activitygen creates a set of trips which is

subsequently used as basis for the routing algorithm

presented in the next section to generate the actual

initial traffic.

4.2.2 Routing

The routing process determines how the OD pairs

defined in the trips are connected. Note that this initial

routing is performed for all generated trips, however

the planned control system can allocate new routes

for autonomous vehicles and cooperating Car2X-

equiped vehicles online once it is integrated. The

routing can be done by representing the road network

as a graph and applying a pathfinding algorithm.

Unlike the line graph used in the road priority

analysis, here the junctions are represented by the

nodes and the roads are modelled by the edges. This

enables a direct connection between the cost and the

associated attributes of the roads. However, a simple

distance or time-based allocation would be unsuitable

here, because in a road network the shortest way

might lead through the city core or residential areas

which is not desirable. Also, such a method would not

consider the actions of other road users and could lead

to smaller roads experiencing more demand than they

could handle while better developed roads might not

use their full capacity during critical traffic situations.

In addition, a traffic scenario shall be simulated that

resembles the real traffic situation in Paderborn,

which therefore must be taken into account during the

routing process. Based on these considerations, the

usage costs for each edge 𝑖 comprise of three cost

components and are recalculated for each trip:

1. Expected travel time (𝐽

,



𝑛



)

2. Road priority assessment (𝐽

,

)

3. Real traffic data (𝐽

,



𝑛



)

The vector 𝑛 contains the number of routes that

includes the respective edge 𝑖 and the vector 𝑟

represents the weighting factors for the cost

components. The resulting costs for crossing edge 𝑖

are therefore:

𝐽





𝑛





𝐽

,



𝑛



𝐽

,

𝐽

,



𝑛





⋅𝑟.

(3)

The expected travel time is a major influencing

factor for drivers’ routing decisions and must be

considered. It can be estimated using the length 𝑙



the considered road section and the expected travel

speed on it. Under ideal circumstances (no other

traffic present) this is equal to the speed limit 𝑣

,

on this road. With increasing traffic demand 𝑛



this road the expected speed decreases which is

modelled by the monotonously falling function

𝑓



𝑛





∈



0,1



and leads to

𝐽

,



𝑛



=𝑙



⋅



𝑣

,

⋅

𝑓



𝑛









(4)

The second cost factor 𝐽

,

is based on the results

of the road priority assessment and depends on the

importance of the road for the network. More

important roads are associated with lower costs while

lower rated roads result in higher costs. In this case, a

linear relation to the priority value was selected.

In order to match the recreated traffic according

to the real traffic measurements, a third cost factor

𝐽

,



𝑛



is introduced for roads with available

measurement data. This factor adds additional costs

to these roads if the current number of routes

containing this road 𝑛



exceeds the number of

vehicles 𝑛

,

recorded by the measurements. Since

the latter value refers to a whole day of

measurements, it is corrected by the share of all

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

374

allocated routes related to the number of all recorded

vehicle crossings on all observed roads resulting in

𝐽

,



𝑛



=max0,𝑛



−𝑛

,

∑





∑



,



.

(5)

The total costs for a route are calculated by adding

up the costs 𝐽





𝑛



of all contained edges 𝑖. To

estimate the best route between the given origin and

destination of a trip Dijkstra’s algorithm is applied to

the graph which is guaranteed to find the connection

with the lowest associated costs. As the initial routing

is performed prior to the actual traffic simulation

there is no need to use a faster but less reliable

algorithm. Also, the order in which the trips are

processed is selected randomly and not

chronologically by their departure time which is

intended to further diversify the routes connecting

different parts of the road network. This way

alternative route options for similar connections are

possible from early on and not after a certain

simulation time to reach the threshold to switch to

another route option (e. g., when the estimated

traveling time-related cost increases for a road section

due to the increased number of routes).

In figure 5, the number of allocated routes for

each road section is depicted on the left side. To better

illustrate the differences on lower frequented roads a

logarithmic colour coding was used. Generally, it can

be seen that Paderborn’s main roads have a greater

number of routes allocated to them while roads in

residential areas or the city centre exhibit much less

demand, which is realistic. To compare the

accordance of the generated routes to the

measurement data, at each available measurement

location the share of realized detections it calculated

(𝑛



/𝑛

,

). The average share for all sensor locations

equals to 100,3% with a variance of 7,3%. An

overview over the deviations of the number of routed

vehicles from the measured data is depicted on the

right side of figure 5. The number of detected and

routed vehicles at the sensor positions are generally

similar although locally the number can deviate

slightly in both directions. A reason for that might be

that not all features of the road network could be

modelled for the Activitygen application, e. g.,

companies with many employees outside dedicated

industrial areas or shopping centres which would

have exceeded the limits of this research.

To show the temporal distribution of the traffic

demand, figure 6 includes the cumulated detection

rates at multiple sensor locations. This is compared to

the cumulated crossing rates resulting from the

generated routes at these positions. There is a good

accordance between both datasets. Both, the morning

and afternoon rush hours are clearly visible and the

values generally match. Slight deviations e. g. at 3:00

and 17:00 are most likely caused by the probability-

based approach for trip generation. Note, that all

crossings registered for a route are allocated to its

departure time on the horizontal axis which explains

the small offset e. g. at 7:00.

Figure 5: Left: Number of routes for each road section (logarithmic) / Right: Deviation routes count from measurements.

Simulation Environment for Trafﬁc Control Systems Targeting Mixed Autonomy Trafﬁc Scenarios

375

Figure 6: Comparison of the measured detection rates and

the generated routes for a whole day at 54 locations.

5 CONCLUSION

In this paper, the modelling of a simulation

environment based on the city of Paderborn for a

future traffic scenario was presented. The simulation

model is built on the software SUMO which handles

the basic vehicle dynamics and is extended by

multiple components. The road network was

imported from OSM, revised manually, analysed, and

converted to a graph representation. Based on that, a

road priority analysis is preformed using different

metrics as well as real traffic data in order to rate the

different road sections’ importance for the whole

system. The results are used in the routing process

and are also useful for the traffic control system

currently in development. To accurately reproduce

the influence of TLS and ensure that they obey the

guidelines and restrictions of the RiLSA, a controller

was designed to implement a given target phase

selected by the control system. Using geometrical

features of the road network, signals, phases, and

additional configuration data were generated

automatically for the TLS. Also, different sensor

types were modelled which support both, stationary

and mobile data collection in order to provide realistic

information to a traffic observer system. To populate

the simulated roads, multiple vehicle types were

created for human-driven and autonomous vehicles.

Based on the road network, statistical and structural

data of Paderborn, trips were generated containing the

desired origin and destination as well as the departure

times of vehicles in the system. Finally, to create

realistic routes, a pathfinding method utilizing a

dynamic cost estimation method was applied.

The next step is the integration of the mentioned

traffic observer to reconstruct a picture of the current

traffic state based on the gathered sensor data. An

observer is currently under development and relies on

a probability-based approach to describe the vehicles’

positions. Key of such a system is the handling of

uncertainty due to incomplete sensor coverage and a

realistic extrapolation of the vehicles’ behaviour. An

in-detail description and evaluation of this system

will be subject for a future publication. Also, the

development and integration of the traffic control

system is due for the future.

ACKNOWLEDGEMENTS

This research was enabled by the Karl-Vossloh-

Stiftung, and we thank them for their support.

REFERENCES

Ahmadzai, F., Rao, K. L., Ulfat, S. (2019). Assessment and

modelling of urban road networks using Integrated

Graph of Natural Road Network (a GIS-based

approach). In Journal of Urban Management, 8(1),

109-125.

Farrag, S. G., Outay, F., Yasar, A. U. H., Janssens, D.,

Kochan, B., Jabeur, N. (2020). Toward the

improvement of traffic incident management systems

using Car2X technologies. In Personal and Ubiquitous

Computing, 25(1), 163-176.

Guo, Y., Ma, J. (2020). Leveraging existing high-

occupancy vehicle lanes for mixed-autonomy traffic

management with emerging connected automated

vehicle applications. In Transportmetrica A: Transport

Science, 16(3), 1375-1399.

Henning, S., Biemelt, P., Abdelgawad, K., Gausemeier, S.,

Evers, H. H., Trächtler, A. (2017). Methodology for

determining critical locations in road networks based on

graph theory. In IFAC-PapersOnLine, 50(1), 7487-

7492.

Krichene, W., Castillo, M. S., Bayen, A. (2016). On social

optimal routing under selfish learning. In IEEE

transactions on control of network systems, 5(1), 479-

488.

Lazar, D. A., Bıyık, E., Sadigh, D., Pedarsani, R. (2021).

Learning how to dynamically route autonomous

vehicles on shared roads. In Transportation research

part C: emerging technologies, 130, 103258.

Lobo, S. C., Neumeier, S., Fernandez, E. M., Facchi, C.

(2020). InTAS--The Ingolstadt Traffic Scenario for

SUMO. arXiv preprint arXiv:2011.11995.

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

Flötteröd. Y. P., Hilbrich, R., Lücken, L., Rummel, J.,

Wagner, P., Wießner, E. (2018). Microscopic traffic

simulation using SUMO. In 2018 21st international

conference on intelligent transportation systems (ITSC)

(pp. 2575-2582). IEEE.

Ma, X., Hu, X., Schramm, D. (2020). Effects of school

activities on traffic flow on the example of traffic

Detections

hou

(x1000)

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

376

scenario from the city Wuhan. In Neue Dimensionen

der Mobilität (pp. 153-163). Springer Gabler,

Wiesbaden.

Maiorov, E. R., Ludan, I. R., Motta, J. D., Saprykin, O. N.

(2019). Developing a microscopic city model in SUMO

simulation system. In Journal of Physics: Conference

Series (Vol. 1368, No. 4, p. 042081). IOP Publishing.

Malena, K., Link, C., Bußemas, L., Gausemeier, S.,

Trächtler, A. (2022). Traffic Estimation and MPC-

Based Traffic Light System Control in Realistic Real-

Time Traffic Environments. In International

Conference on Vehicle Technology and Intelligent

Transport Systems, International Conference on Smart

Cities and Green ICT Systems (pp. 232-254). Springer,

Cham.

Malena, K., Link, C., Mertin, S., Gausemeier, S., Trächtler,

A. (2021a). Online State Estimation for Microscopic

Traffic Simulations using Multiple Data Sources. In 7th

International Conference on Vehicle Technology and

Intelligent Transport Systems (pp. 386-395).

Malena, K., Link, C., Mertin, S., Gausemeier, S., Trächtler,

A. (2021b). Validation of an Online State Estimation

Concept for Microscopic Traffic Simulations. In 2021

IEEE Transportation Electrification Conference &

Expo (ITEC) (pp. 212-217). IEEE

Mertin, S., Malena, K., Link, C., Gausemeier, S., Trächtler,

A. (2020). Macroscopic Traffic Flow Control using

Consensus Algorithms. In 2020 IEEE 23rd

International Conference on Intelligent Transportation

Systems (ITSC) (pp. 1-8). IEEE.

OpenStreetMap contributors (2022). Paderborn Map, from

https://www.openstreetmap.org.

Forschungsgesellschaft für Straßen- und Verkehrswesen

(2015). Richtlinien für Lichtsignalanlagen. RiLSA:

Lichtzeichenanlagen für den Straßenverkehr. FGSV R1

- Regelwerke, FGSV-321, Köln

Schweizer, J., Rupi, F., Filippi, F., Poliziani, C. (2018).

Generating activity based, multi-modal travel demand

for SUMO. In EPiC Series in Engineering, 2, 118-133.

Simulation Environment for Trafﬁc Control Systems Targeting Mixed Autonomy Trafﬁc Scenarios

377