Urban Trafﬁc Incident Detection for Organic Trafﬁc Control:

A Density-based Clustering Approach

Ingo Thomsen, Yannick Zapfe and Sven Tomforde

Intelligent Systems, Christian-Albrechts-Universit

at zu Kiel, 24118 Kiel, Germany

Keywords:

Organic Trafﬁc Control, Trafﬁc Flow Analysis, Trafﬁc Incident Detection, Trafﬁc Management.

Abstract:

The trafﬁc demands in urban road networks can ﬂuctuate immensely. The Organic Trafﬁc Control (OTC) of-

fers a resilient trafﬁc management to control such trafﬁc demands. An additional challenge is the detection of

unforeseen trafﬁc incidents. To enhance the capabilities of OTC accordingly, we outline a trafﬁc incident algo-

rithm based on DBSCAN, a density-based clustering algorithm: In a simulated urban road network, equipped

with trafﬁc light controllers at intersections, vehicle detectors are used to gather trafﬁc ﬂow data. The cluster-

ing of this time series data to detect simulated road blockages is expanded using various ﬁlters. This extension

of the initial clustering is the result of an manual evaluation process, which shows the principal applicability

of this approach.

1 INTRODUCTION

Increasing mobility and consequential rising trafﬁc

demands elevate trafﬁc density on the streets, espe-

cially in urban areas. Therefore, the average travel

times for individual trafﬁc participants can rise. As

trafﬁc situations are in constant change at a spatial

and temporal level, trafﬁc management solutions re-

quire mechanisms for adaptation at runtime.

The Organic Trafﬁc Control (OTC) system (Som-

mer et al., 2016) is self-organised trafﬁc management

system. Each intersection in a network is controlled

by an instance of the OTC, which decides locally for

each node if and how the trafﬁc light signalisation

is to be adapted. For this, the trafﬁc conditions are

constantly analysed, and the decision behaviour is im-

proved over time. This facilitates resilient trafﬁc man-

agement.

In addition to the common ﬂuctuation in trafﬁc

demands, various trafﬁc incidents – accidents, road

blockades, unscheduled maintenance, or construction

work – can have a severe impact on the trafﬁc be-

haviour. To boost the adaptive capacities of the OTC

system in the context of urban road networks, inci-

dent detection capabilities are highly beneﬁcial for

more efﬁcient and accurate trafﬁc management. This

paper explores the possibility to detect incidents in

urban road networks by employing a density-based

clustering algorithm (DBSCAN) (Ester et al., 1996)

for trafﬁc incident detection. An intermediate goal is

to equip the OTC with this capability. In contrast to

state-of-the-art approaches from highways, the inci-

dent detection problem in urban areas is more com-

plex due to a more heterogeneous trafﬁc model (e.g.,

intersections with trafﬁc lights, more distributed traf-

ﬁc demands, primary and secondary roads) and par-

ticipant behaviour (e.g., unloading lorries, stopping

busses). This can introduce patterns in the detector

data that are very similar to those of actual incidents.

The remainder of this paper is organised as fol-

lows: The next section provides information on the

Organic Trafﬁc Control and trafﬁc incidents detec-

tion. Section 3 then describes underlying assump-

tions about the trafﬁc model under consideration. The

next two sections outline the main contribution of this

work: In Section 4 the proposed incident detection is

presented and Section 5 describes the corresponding

evaluation, which includes experimental setup, eval-

uation, and results. As an outlook, Section 6 pro-

poses the next steps towards an incident-aware self-

organised trafﬁc control behaviour based on OTC. Fi-

nally, Section 7 summarises the paper.

2 BACKGROUND

As a basis for the presented approach, this section in-

troduces the OTC system and brieﬂy summarises the

state-of-the-art in incident detection.

152

Thomsen, I., Zapfe, Y. and Tomforde, S.

Urban Trafﬁc Incident Detection for Organic Trafﬁc Control: A Density-based Clustering Approach.

DOI: 10.5220/0010454101520160

In Proceedings of the 7th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2021), pages 152-160

ISBN: 978-989-758-513-5

2.1 Organic Trafﬁc Control

The Organic Trafﬁc Control (OTC) system is a traf-

ﬁc management system, developed according to the

principles of Organic Computing (M

uller-Schloer and

Tomforde, 2017). It is a self-adaptive and self-

organising (SASO) system, which distributes the

complexity of decision-making processes among au-

tonomous agents, which in turn can cooperate with

each other. Such SASO systems tackle the complex-

ity of controlling runtime behaviour by distributing

the decision-making processes among autonomous

agents, which in turn cooperate with each other:

Goals can be reached in a large system without

centralised ruling. It consists of a productive part

and a control mechanism, based on a multilevel ob-

server/controller architecture by (Tomforde et al.,

2011). This architecture is outlined in Figure 1 and

consists of 3 levels:

Level 0. Here, the the System under Observation

and Control (SuOC) is situated – the (simulated) traf-

ﬁc system equipped with sensors (detector) and traf-

ﬁc light controllers (actuator). The signal settings of

these controllers can be altered.

Level 1. This level contains an observer that pro-

cesses the sensor data from the SuOC to create a

model of the current trafﬁc (ﬂow) situation. This

is handed over to the Controller, which is based on

a Learning Classiﬁer System (LCS). The controller

Figure 1: Overview of the multilevel OTC architecture.

then selects a ﬁtting rule to alter the trafﬁc light sig-

nalisation in the SuOC This is the “online” layer.

Level 2. This is the “ofﬂine” layer, which is acti-

vated, in case no appropriate rules can be found. It

uses an evolutionary algorithm to generate new rules

for Layer 1. These newly found rules are evaluated

using a trafﬁc simulator (Aimsun SLU, 2020).

Each intersection is equipped with such an OTC

control module and, through interaction. additional

functionality can be provided, e.g., dynamic route

guidance (DRG) or progressive signal systems (PPS),

also known as “green waves” in trafﬁc.

2.2 Trafﬁc Incident Detection

Trafﬁc control has to continuously evaluate informa-

tion about the current road situation, including traf-

ﬁc incidents. The California Algorithm (Payne and

Tignor, 1978), the incident detection approach imple-

mented in OTC, is a decision tree algorithm based on

four states: incident-free, incident termination, ini-

tial detection, and incident continuation, all based on

occupancy values at several locations. This method

focuses mainly on highways and is not necessarily

equally adequate for urban areas, as argued by (Som-

mer et al., 2016). For further reading, (Parkany and

Xie, 2005) presents an exhaustive list of incident de-

tection methods for highways and arterial roads.

The information on the trafﬁc ﬂow is usually rep-

resented as time series data. A typical part of the cor-

responding preprocessing is the detection of anoma-

lies, alongside the related, but distinct novelty and

outlier detection. (Pimentel et al., 2014) describe

anomaly detection as recognising that test data differs

signiﬁcantly from training data. They also outline dif-

ferent approaches, based on information theory, do-

main, reconstruction, probability, and distance. In this

context anomalies are localised trafﬁc incidents that

deviate considerably from normal trafﬁc behaviour

and have a signiﬁcant inﬂuence on the overall net-

work performance. For instance, such incidents could

be road works or accidents (see Section 3).

Cluster analysis, the identiﬁcation of groups of

similar points in (time series) data, is a common tech-

nique for statistical data analysis, with numerous al-

gorithms based on connectivity, centroids, grid, distri-

bution, and density. Such a well density-based clus-

tering algorithm, DBSCAN (Ester et al., 1996), is

used as a base for incident detection as described in

more detail in Section 4.

Urban Trafﬁc Incident Detection for Organic Trafﬁc Control: A Density-based Clustering Approach

153

3 MODEL ASSUMPTIONS

The approach for detecting trafﬁc incidents described

in this paper is intended to be used in the context of

Organic Trafﬁc Control (see Section 2.1). Therefore,

this section gives a summary of the broader underly-

ing domain model and the simulation setup.

Urban Road Networks. In contrast to highways or

arterial roads, urban road networks do not have a priv-

ileged “main road” with a dominating trafﬁc load.

Single or multi-lane sections with speed limits con-

nect intersections, which are equipped with one traf-

ﬁc light controller (TLC) each to operate trafﬁc lights

at each incoming section. The signalisation phases of

the TLCs are displayed periodically based on a con-

trol cycle. The incoming and outgoing sections also

have detectors for counting vehicles (such as induc-

tion loops embedded in the street surface), thus al-

lowing to calculate the trafﬁc ﬂow for each section.

The network may include junctions where no roads

“intersect” and which are not controlled by a TLC.

To facilitate other OTC features (PSS or DRG), all

incoming roads of an intersection are equipped with

variable message signs (VMS) to relay information to

the drivers. Additionally, the junction controllers may

communicate with immediate neighbours.

Trafﬁc Demand. The occurring trafﬁc load can be

deﬁned as an origin-destination (O/D) matrix which

deﬁnes how many vehicles traverse from each origin

to each destination per hour. An O/D matrix may also

be time-variant to model changing demands (e.g., dur-

ing a simulation).

Incidents. At this stage, a general incident model is

applied. Real-life events and their immediate effect

are represented as a full roadblock. They are located

within a section with ﬁxed start and end times. Sec-

tion 6 suggests a prospective, more complex model.

4 DETECTION APPROACH

The intention of this work is to evaluate DBSCAN

with respect to identifying incidents based on prepro-

cessed trafﬁc ﬂow values. As a result of the evalu-

ation process outlined in Section 5.7, detection im-

provements (ﬂattening and domain knowledge) were

added to this approach and are described in this sec-

tion.

4.1 DBSCAN

Density Based Spatial Clustering of Applications with

Noise (DBSCAN) (Ester et al., 1996) is an established

algorithm for clustering and anomaly detection. It is

efﬁcient for large amounts of data and does not re-

quire domain knowledge. Also, the number of target

clusters does not have to be speciﬁed beforehand and

the clusters can be of arbitrary shape.

The key idea of DBSCAN is to form clusters of

core objects with a minimum neighbourhood density

and further objects that are connected to these cores.

The neighbourhood density is determined by the num-

ber of other objects surrounding a data object within

a given radius. Other objects that are not connected

with objects belonging to a cluster are considered as

noise. Here, the objects are the points of the trafﬁc

ﬂow data. With respect to a distance measure dist and

the two algorithm parameters ε and minPts, the fol-

lowing concepts are used:

ε-neighbourhood. This is the set of points N

that

is located near a point p within a given radius ε:

(p) = {p

| dist(p, p

) ≤ ε} (1)

Directly Density-reachable. All points p

are di-

rectly density-reachable from a point p if the follow-

ing applies:

∈ N

(p) (2)

(p)| ≥ MinPts (3)

Density-reachable. All points p

are density-

reachable from a point p if a chain of points p =

,..., p

= p

exists such that p

and p

i+1

are directly

density-reachable for each i ∈ {1, ..., n}.

Density-connected. Two points p and p

are

density-connected if a point q exists from which both

are density-reachable.

Cluster. The set of points C for which applies:

∀p, p

: if p

density-reachable from p ∈ C (4)

∀p, p

∈ C : p is density-connected to p

(5)

Noise. All points that do not belong to any cluster

are categorised as Noise.

An important part of DBSCAN is the choice of

the distance measurement for determining the ε-

neighbourhood. Section 5.5 lists the measures used

for this work.

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

154

(a) Control cycle starting at 0s. (b) Control cycle starting at 90s. (c) Control cycle starting at 180s.

Figure 2: Trafﬁc ﬂow at intersection labelled “2;1 1;1” (see Figure 3). Flattened and unﬂattened time series show the ﬂows

in vehicles/hour – indicated on the y-axis – for the second 45s of each of the ﬁrst three control cycles at the beginning of a

simulation. Blue curve: no ﬂattening; yellow: ﬂattening weight of 66 %; purple: 50 %; green: 33 %.

4.2 Flattening

The raw detector values are preprocessed to form time

series of ﬂow values – one data point per section and

control cycle. In addition, the time series are ﬂattened

(dampened) using a weighted average. Here, previ-

ous trafﬁc ﬂows are taken into account with more re-

cent values having more inﬂuence. The weight w is

the only parameter, denoting the percentage of a data

point in the weighted average. For a ﬂow value time

series x

in section s at the starting time t with a con-

trol cycle duration t

, the ﬂattening f is calculated re-

cursively:

f (x

) =

(

if t = 0

∗

100

+ (1 −

100

) ∗ f (x

t−t

) otherwise

(6)

For instance, Figure 2 shows the time series of

three cycles at the beginning of a simulation in the 2x2

Manhattan grid (see Section 5.1) for several weights:

For the ﬁrst cycle, there is no “previous” data, so the

ﬂattened times series are the same. This is then taken

into calculation for the second cycle and ﬂattens the

time series accordingly. In the last chart, both previ-

ous time series are integrated into the weighted aver-

age. E.g., for the ﬂattening weight of 50, the current

time series is accountable for 50 % of the weighted

average; the previous time series for 50 % of the re-

maining 50, resulting in a 25 % share for the ﬁrst time

series. It shows the intended effect of ﬂattening the

curve: Smaller and fewer spikes.

4.3 Domain Knowledge

The integration of knowledge about the model can

help to improve the detection. It can be applied for

initial detection (to increase the detection rate) or vali-

dation (to reduce the false positives). During the eval-

uation process described in Section 5.7, several ob-

servations about the trafﬁc networks, demands, and

incidents were utilised.

5 EVALUATION

To assess the performance of the incident detection

conceived in Section 4, the algorithm was executed

using varying parameter sets on trafﬁc ﬂow data ob-

tained from trafﬁc simulations. This section describes

the experimental setup with respect to the model from

Section 3 followed by the initial algorithm parameters

and a description of the evaluation process.

5.1 Road Networks

The trafﬁc simulation was based on two regular, fully

connected Manhattan grids as the initial urban road

networks (see Figure 3). Only cars as vehicles were

considered. All sections had a length 150m – roughly

the section length of Barcelona’s chequered inner-city

road network – with one lane plus additional side

lanes for left-turns.

Urban Trafﬁc Incident Detection for Organic Trafﬁc Control: A Density-based Clustering Approach

155

Following experiments of earlier work by (Tom-

forde et al., 2008), each intersection was controlled by

an identical ﬁxed-time controller (FTC) using 4 signal

phases (with a transition time of 5s in between) and a

control cycle time of 90s. In the phase I, cars on hor-

izontal sections (or leaving them by turning right) are

served for 25s. The remaining cars turning left are

served in the phase II for 10s. The vertical trafﬁc is

handled analogously in phases III and IV.

5.2 Trafﬁc Demand

The trafﬁc demands in Table 1 were also adopted from

(Tomforde et al., 2008), and can be categorised in

primary, secondary, and tertiary demands along with

the “straight” O/D connections. For the 2x2 grid, the

most heavily used O/D pairs are A to B and D to C,

with decreasing demand in their opposite directions

and the vertical straight connections. In the case of

the 3x3 grid, the demands as in Table 2 were similar:

The horizontal directions carry the heavier trafﬁc load

and direction A to B the heaviest.

(a) 2x2 Manhattan grid

(b) 3x3 Manhattan grid

Figure 3: Manhattan network with 4-phased FTCs.

The described scenario had a total duration of 1:15

hours with the ﬁrst 15 minutes being the warm-up

phase, which was not considered by the incident de-

tection algorithm.

Table 1: Trafﬁc demands in veh/h for 2x2 Manhattan grid.

O/D demand opposite direction

A – B 400 200

C – D 200 400

E – F 150 150

G – H 150 150

others 10 10

Table 2: Trafﬁc demands in veh/h for 3x3 Manhattan grid.

O/D demand opposite direction

A – B 400 200

C – D 300 150

E – F 300 150

G – H 150 150

I – J 150 150

K – L 150 150

other 10 10

5.3 Incidents

All incidents start at 00:45 and last half an hour un-

til the simulation end. This is the second half of

the simulation phase after the warm-up. In addition

to the incident-free simulation run, several scenarios

with one incident were created. Depending on the in-

cident placement along the varying trafﬁc demands

described above, these scenarios fall into three cate-

gories. Due to symmetries of the topology and de-

mands, one incident each was chosen from each cate-

gory.

5.4 Simulation

For the simulation of road networks, incidents, and

trafﬁc demands (and the actual routing of the cars), a

commercial trafﬁc modelling and simulation software

was used: AIMSUN Next (Aimsun SLU, 2020) – an

obvious choice, as it is used for OTC as well (see Sec-

tion 2.1). A client was programmed that used the sim-

ulator’s API to extract vehicle counts from a running

experiment. In this client-server setup, the incident

detection, as well as the generation of plots and log

ﬁles for the evaluation were implemented as part of

the server.

5.5 Distance Measures

The applied distance measures for DBSCAN are

brieﬂy explained below. A point of the input data set

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

156

consists of a time series of ﬂow values for a certain

section over the course of a single control cycle – one

measurement per second. DBSCAN is executed once

per section and only time series of the same section

are compared using the distance measure.

Euclidean Distance. The simple length of a straight

line between two points in Euclidean space is widely

used and as the time series have the same resolution,

the exact same points in time are compared.

Dynamic Time Warping (DTW). This similarity

measure by (Berndt and Clifford, 1994) searches a

so-called warping path for two time series: A con-

tinuous and monotonous sequence of pairs of points

from both series. This path has to cover both time se-

ries completely. The cumulative distance between the

data points is calculated based on another measure. In

this work, Euclidean distance is used and the resulting

DTW distance is normalised by the path length.

Figure 4 depicts an artiﬁcial example of both dis-

tance measures: While the Euclidean distance only

matches points at the same time, points from one se-

ries can be matched with multiple points from the

other series in DTW.

(a) Euclidean

(b) Dynamic Time Warping

Figure 4: Artiﬁcial example for applying distance measures

between the two time series – red and blue – which are con-

nected by multiple lines that show the different data point

matching for DTW (b) and Euclidean distance (a).

Polynomial Approximation. Two time series are

approximated by polynomial functions of a certain

degree – linear and cubic in this work. Then the Eu-

clidean distance between the two lists of polynomial

coefﬁcients is calculated.

Average Distance. The absolute difference be-

tween the average values of two time series.

Relative Distance. This is a derived measure: One

of the above measures is used to calculate a distance

which is then normalised using the maximum of the

two time series averages. Therefore, the distance is

relative to the larger of two average ﬂow values.

5.6 Criteria

To evaluate the choice and combinations of detection

algorithm parameters, these three criteria of decreas-

ing priority were considered:

1. Detection Rate: The most important criterion to

be maximised is similar to the true positive rate in

binary classiﬁcation tasks. Here, a correct detec-

tion of an incident includes discerning the section

where it occurs. Obviously, the incident should

only be detected after the incident has actually

started. As each scenario contains either one or no

incident, the detection rate is the ratio of correctly

identiﬁed and expected incidents for all scenarios

in a model.

2. False Alarms: Number of identiﬁed incidents that

do not correspond to a simulated incident (ei-

ther non-existent or located in a different section).

This criterion should be minimal. It is prioritised

lower than the detection rate: Some false alarms

may be tolerated to achieve better detection.

3. Detection Delay: This has to be considered when

applying ﬂattening (see Section 4.2), as incidents

then have to persist multiple control cycles to be-

come signiﬁcant for the distance measure. Obvi-

ously, smaller time delays are preferable.

5.7 Evaluation Process

The concrete evaluation leading to the results in Sec-

tion 5.8 was an iterative process which comprises of

several phases: Initially, a rather generic set of pa-

rameter combinations was used to prove the general

applicability. This was followed by a manual opti-

misation of the algorithm parameters. Dependencies

were identiﬁed along with parameters that have little

to no impact, which can then be locked at a certain

value. Equally, a parameter can already have an obvi-

ous optimal value that outperforms others, so it can be

locked at that optimum as well. These values were the

base for ﬁnding a good combination through manual

exploration based on the 2x2 Manhattan grid, which

were then evaluated for the 3x3 grid as well. The goal

Urban Trafﬁc Incident Detection for Organic Trafﬁc Control: A Density-based Clustering Approach

157

was to ﬁnd a parameter set that is optimal according

to evaluation criteria above for both models.

Initial Parameter Combinations. In their original

description of DBSCAN, (Ester et al., 1996) state that

minPts does not have a signiﬁcant inﬂuence and can

be set to 4 as default. This was done for the initial val-

idation. To ﬁnd candidates for the more important ε

parameter, k-distance plots for runs with and without

an incident were inspected. Depending on distance

measure and ﬂattening, the ε values were chosen. For

instance, the values 500, 400, 300, 200, and 100 were

selected for “Euclidean distance without ﬂattening”.

For ﬂattening, the weights 33, 50, 66 and 0 were se-

lected. Together with 5 distance measures (all except

the relative distance), this resulted in 100 initial pa-

rameter combinations.

Only moderate detection rates were achieved, cou-

pled with high numbers of false alarms. This indi-

cated the need for adjustments of the incident detec-

tion.

Incident Validation. To improve the discrimination

of trafﬁc ﬂuctuation against real incidents, domain

knowledge was applied: While decreased trafﬁc is ex-

pected in the “incident section” and its downstream

sections, the load should increase upstream as well as

in “detour sections”. A validation mechanism based

on this was devised, which decreased the number of

false alarms signiﬁcantly, while the detection rates

were conserved.

Finding Incident Indicators. To identify the rea-

sons for weaker detection rates, the correctly detected

incidents were analysed in more detail. These were

mostly associated with high trafﬁc demand. To bet-

ter detect secondary and tertiary incidents, the rel-

ative distance measure was introduced. After test-

ing 10.000 parameter combinations, it showed to be

mainly suitable for the incident indicators rather than

for validators. From there on, the relative distance

was only used for indicators, while the other measures

were only applied for validators.

Detection Delay. A deeper look into the detection

delay showed, that heavier ﬂattening leads to longer

delay: All optimal combinations with regard to detec-

tion rate and false alarms included ﬂattening, which

led to signiﬁcant delays of at least two control cycle

periods.

Incident Filters. Incidents with certain properties

common to false alarms were identiﬁed. One ﬁnd-

ing was that most of them were situated on the edge

of the road model, possibly due to the way AIMSUN

Next simulates the incoming trafﬁc. These were ﬁl-

tered out.

Additionally, false alarms were discovered almost

exclusively in scenarios that also included incidents.

It showed that some validators conﬁrmed too many

other incidents, which resulted in false alarms. There-

fore, they were restricted to validate at most one inci-

dent.

minPts. Prior to the last optimisation step, the so-

far locked parameter minPts was investigated once

more. The values 2, 3, 4, 5, 6 and 8 were tested for

all sets of distance measures, ε values and ﬂattening

weights as well as for ﬁlters, indicator and valida-

tor instances: Especially good combinations of dis-

tance measures and ε work equally well for all tested

minPts values, but as no signiﬁcant changes occurred,

minPts = 4 was retained.

Parameter Optimisation. Finally, the optimal re-

sults for the 2x2 grid were evaluated in the 3x3 grid in

order to optimise the parameters for unﬂattened data.

The optimal parameter sets according to detection rate

and false alarm count are chosen from those combina-

tions that achieved an average detection delay of less

than two control cycle durations (180 s).

5.8 Results

Table 3 presents the results for the 2x2 Manhat-

tan grid. It shows that the incident ﬁlter reduces

the number of false alarms for all distance mea-

sures. Although the results have been optimal for

some combinations even before introducing the ﬁl-

ter, more combinations of different parametrisation

now show nearly optimal results. An optimal result

was achieved parametrisation when using average dis-

tance for validation and relative distance for indica-

tion.

The Table 4 shows the best ε values for each com-

bination of indicator and validator distance measure

with unﬂattened data in the 3x3 Manhattan grid. The

results indicate that in the case of unﬂattened data no

dependency between the indicator and validator pa-

rameter sets exists. For all validator distance mea-

sures, the same ε values work best for both relative

distance measures. In return, the relative average dis-

tance with ε set to 0.95 works best for all validator

distance measures and especially outperforms the rel-

ative DTW with respect to false alarms. The validator

distance measures do not differ signiﬁcantly in their

results, but the Euclidean distance measure with ε set

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

158

Table 3: Best parametrisations after introducing the incident

ﬁlter from initial parameter combinations in a 2x2 Manhat-

tan grid.

Validation Indication

Relative detec- false

distance ε distance ε tion alarms

measure measure rate

Euclidean 200 Average 0.8 1 0

DTW 12 Average 0.8 1 0

Linear 20 Average 0.8 1 0

Cubic 24 Average 0.8 1 0

Average 8 Average 1 1 0

Euclidean 200 DTW 0.8 1 0

DTW 18 DTW 0.8 1 0

Linear 20 DTW 0.8 1 1

Cubic 12 DTW 0.8 1 1

Average 16 DTW 0.8 1 0

to 150 and the cubic approximation distance with 9

as ε achieve slightly better results than the other mea-

sures.

Table 4: Best parametrisations after introducing the inci-

dent ﬁlter from promising parameter combinations in a 3x3

Manhattan grid.

Validation Indication

Relative detec- false

distance ε distance ε tion alarms

measure measure rate

Euclidean 150 Average 0.95 1 4

DTW 9 Average 0.95 1 7

Linear 10 Average 0.95 1 7

Cubic 9 Average 0.95 1 5

Average 8 Average 0.95 1 7

Euclidean 150 DTW 0.8 1 20

DTW 9 DTW 0.8 1 24

Linear 10 DTW 0.8 1 24

Cubic 9 DTW 0.8 1 23

Average 8 DTW 0.8 1 26

6 FUTURE WORK

The evaluation process in Section 5 showed that, in

principle, DBSCAN combined with some enhance-

ments can be employed for incident detection in sim-

ple urban road networks. Several assumptions are

made in the context of the experiments, and future

research could address the consequent limitations.

Model Expansion. More realistic – and complex

– road networks and trafﬁc demands have to be ad-

dressed, e.g., real-life networks with challenging traf-

ﬁc loads like during rush-hour. Also, the underlying

incident model can be expanded dramatically. Instead

of a “binary” roadblock, real-life incidents can have

varying effects with respect to spatial (location, mov-

ability, extend, . . . ) or temporal characteristics (dura-

tion, regularity, . . . ). Finally, an incident may occur

isolated: Different types of trafﬁc incidents may ap-

pear simultaneously or with relatively small delay.

Algorithm Enhancements. The presented incident

detection can be extended: Additional distance mea-

sures or even a different clustering algorithm are con-

ceivable, e.g., the Local Outlier Factor (LOF) by

(Breunig et al., 2000), which shares similarities with

DBSCAN. Also, an expansion of the model might

provide domain knowledge, which could be beneﬁ-

cial for incident indication and validation.

Parameter Optimisation. The evaluation in Sec-

tion 5.7 for a restricted selection of algorithm param-

eters already generated thousands of test runs. To-

gether with the extensions proposed above, a manual

inspection of parameters sets is no longer feasible.

Hyperparameter optimisation from machine learn-

ing could provide suitable techniques, for instance,

grid search, evolutionary algorithms, Bayesian or

gradient-based optimisation

Integration with OTC. The structured exploration

of parameters on potentially much more complex

models calls for an automated system for setting up

and conducting experiments as well as the subsequent

evaluation. Consequently, the integration into the Or-

ganic Trafﬁc Control could improve the self-adaptive

and self-organising capabilities of the OTC.

7 CONCLUSION

In this work, a trafﬁc incident detection approach

in the context of urban road networks is presented,

which employs the density-based clustering algorithm

DBSCAN together with some enhancements. These

include incident validation using surrounding sec-

tions, a combination of different algorithm parameter

sets for the detection and validation of incidents, and

ﬁltering out incidents with characteristics common to

false alarms. These enhancements are results of the

evaluation process, which eventually show the princi-

pal applicability of the proposed approach.

Urban Trafﬁc Incident Detection for Organic Trafﬁc Control: A Density-based Clustering Approach

159

The experimental setup is limited compared to

real-life networks, trafﬁc demands and incidents. To-

gether with the numerous and often manual evaluation

steps, further research steps can be identiﬁed, which

were also presented as future work.

REFERENCES

Aimsun SLU (2020). Aimsun Next Professional, Version 20.

Barcelona, Spain.

Berndt, D. J. and Clifford, J., editors (1994). Using dy-

namic time warping to ﬁnd patterns in time series,

volume 10. Seattle, WA, USA.

Breunig, M. M., Kriegel, H.-P., Ng, R. T., and Sander, J.

(2000). Lof. In Dunham, M., Naughton, J. F., Chen,

W., and Koudas, N., editors, Proceedings of the 2000

ACM SIGMOD international conference on Manage-

ment of data - SIGMOD ’00, pages 93–104, New

York, New York, USA. ACM Press.

Ester, M., Kriegel, H.-P., Sander, J., and Xu, X. (1996).

A density-based algorithm for discovering clusters in

large spatial databases with noise. pages 226–231.

AAAI Press.

uller-Schloer, C. and Tomforde, S. (2017). Organic Com-

puting – Technical Systems for Survival in the Real

World. Autonomic Systems. Birkh

auser. ISBN: 978-

3-319-68476-5.

Parkany, E. and Xie, C. (2005). A complete review of inci-

dent detection algorithms & their deployment: What

works and what doesn’t.

Payne, H. J. and Tignor, S. C. (1978). Freeway incident-

detection algorithms based on decision trees with

states. In Urban system operation and freeways,

Transportation research record. National Academy of

Sciences, Washington, DC.

Pimentel, M. A., Clifton, D. A., Clifton, L., and Tarassenko,

L. (2014). A review of novelty detection. Signal Pro-

cessing, 99:215–249.

Sommer, M., Tomforde, S., and H

ahner, J. (2016). An or-

ganic computing approach to resilient trafﬁc manage-

ment. In McCluskey, T. L., Kotsialos, A., M

uller, J. P.,

ugl, F., Rana, O., and Schumann, R., editors, Auto-

nomic Road Transport Support Systems, pages 113–

130. Birkh

auser, Basel.

Tomforde, S., Prothmann, H., Branke, J., H

ahner, J., Mnif,

M., M

uller-Schloer, C., Richter, U., and Schmeck, H.

(2011). Observation and control of organic systems.

In Organic Computing—A Paradigm Shift for Com-

plex Systems, pages 325–338. Springer.

Tomforde, S., Prothmann, H., Rochner, F., Branke, J.,

ahner, J., M

uller-Schloer, C., and Schmeck, H.

(2008). Decentralised progressive signal systems

for organic trafﬁc control. In 2008 Second IEEE

International Conference on Self-Adaptive and Self-

Organizing Systems, pages 413–422. IEEE.

VEHITS 2021 - 7th International Conference on Vehicle Technology and Intelligent Transport Systems

160