Approaches to Automatic Road Trafﬁc Incident Detection and Incident

Forecasting

oren Striewski

, Ingo Thomsen

and Sven Tomforde

Intelligent Systems, Kiel University, 24118 Kiel, Germany

Keywords:

Automatic Trafﬁc Incident Detection, Trafﬁc Management, Adaptive Trafﬁc Control.

Abstract:

Traditional trafﬁc light controllers are unable to respond to variations in trafﬁc demand as they generally rely

on ﬁxed-time signalisation with predeﬁned sequences. This work presents two algorithms, one for incident de-

tection and one for congestion forecasting. The Extended California Algorithm (ECA), an incident detection

algorithm, addresses ﬂaws in the established California Algorithm. The congestion forecast algorithm detects

occurrences when trafﬁc exceeds the capacity of the accessible roads by comparing the present dynamic road

capacity with the anticipated future trafﬁc ﬂow. Both are then compared with the established California algo-

rithm.

1 INTRODUCTION

Urban road networks are characterised by signalised

intersections and the general lack of highly prioritised

roads. With rising demands, trafﬁc management so-

lutions must make use of the current infrastructures

while enhancing signalisation. Since common trafﬁc

light controllers (TLC) rely on ﬁxed-time signalisa-

tion with predeﬁned sequences, they cannot adapt to

variations in the volume of trafﬁc.

Intelligent Trafﬁc Management systems such as

the Organic Trafﬁc Control system (OTC) (Tomforde,

2012) seek to maximise the trafﬁc ﬂow by chang-

ing the signalisation at run-time with respect to the

current trafﬁc conditions. As OTC is founded on

the design principles of Organic Computing (Proth-

mann et al., 2011b), it has the prerequisites to be

self-organised: It can handle unexpected trafﬁc sce-

narios while adapting to changes in its surroundings

(e.g. accidents). Previously, the OTC system was en-

hanced with a forecasting component to predict fu-

ture trafﬁc patterns (see (Sommer et al., 2013)). With

these projections, trafﬁc jams can be anticipated and,

in the long run, avoided. The beneﬁts of signalisation

adaptation over merely detecting congestions include

faster reaction times, shorter travel durations, fewer

stopping vehicles, and reduced pollutant emissions.

Additionally, trafﬁc congestion also causes stress to

https://orcid.org/0000-0002-5773-9266

https://orcid.org/0000-0002-0850-4786

https://orcid.org/0000-0002-5825-8915

drivers and increases the likelihood of accidents, ac-

cording to (Marchesini and Weijermars, 2010).

This work introduces two algorithms, one for inci-

dent detection and the other for congestion forecast-

ing. The Extended California Algorithm (ECA), an

incident detection algorithm, addresses ﬂaws in the

original California Algorithms. This congestion fore-

cast algorithm evaluates occurrences when trafﬁc ex-

ceeds the capacity of accessible roads by comparing

the present dynamic road capacity with the antici-

pated future trafﬁc ﬂow.

The remainder of this paper is structured as fol-

lows. Section 2 gives a brief overview of the ﬁeld

of incident detection and forecasting. Section 3 in-

troduces the architecture of the OTC system and its

Automatic Incident Detection component. Section 4

explains the Enhanced California Algorithm. The in-

cident forecast algorithm is described in Section 5 and

Section 6 reports the experimental results. Section 7

concludes with a summary of the work presented and

outlines further research.

2 BACKGROUND

2.1 Incident Detection

Incident detection systems can be divided into many

categories, according to (Guin, 2004) and (Parkany

and Xie, 2005), who also evaluated and compared

the results of the various algorithms. The categories

Striewski, S., Thomsen, I. and Tomforde, S.

Approaches to Automatic Road Trafﬁc Incident Detection and Incident Forecasting.

DOI: 10.5220/0011859200003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 289-296

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)

289

of smoothing algorithms, artiﬁcial intelligence algo-

rithms, and image processing algorithms are men-

tioned in addition to those identiﬁed by Guin. In his

thorough review, (ElSahly and Abdelfatah, 2022) lists

examples from these categories.

Comparative algorithms assess tracked trafﬁc

metrics (such as volume or speed) in relation to prede-

termined thresholds. This group includes the 10 Cali-

fornia Algorithms (Payne and Tignor, 1978), of which

number 7 and 8 are the ones most applied and used

for comparison. Statistical algorithms spot deviations

from the norm in trafﬁc patterns by employing statis-

tical methods. The measured trafﬁc data are treated as

time series and are compared to predicted or histori-

cal data for deviations which might indicate incidents.

These trafﬁc theory-based algorithms include e.g. the

McMaster algorithm (A.I. and Hall, 1989). Another

representative is the All-Purpose Incident Detection

(APID) (Masters et al., 1991). As extension of the

California Algorithm 7 it distinguishes between low,

medium, and high volume trafﬁc and checks for com-

pression waves and incident persistence.

Another summary (Rao and Rao, 2012) includes

developments in measuring urban trafﬁc congestions

globally and establishes two primary variables inﬂu-

encing trafﬁc congestion: micro-level (for example,

the phenomena of too many people wanting to travel

on the same road at the same time) and macro-level

factors (relating to the overall demand of road usage

such as land-use patterns or regional economic dy-

namics). The analysis revealed that there are numer-

ous alternative methods for identifying trafﬁc jams in

urban areas. Rao also provides a summary of the com-

mon congestion measuring measures, such as speed,

trip time, delay, and volume, including counter criti-

cism of each of one of those.

2.2 Incident Forecasting

Kurihara (Kurihara, 2013) proposes an approach

based on ant colony optimisation. It uses a model of

ant behaviour and their use of pheromones for com-

munication. In order to calculate and forecast short-

term trafﬁc congestions at one-minute intervals, in-

tersection computers (also known as “road agents”)

collect measured trafﬁc ﬂows from locally positioned

sensors. The the local trafﬁc ﬂow density is calcu-

lated and pheromones for forecasting the congestion

as well as the density are sent to neighbouring road

agents. A simulation based on a Manhattan-style road

network lead to higher accuracy in congestion fore-

casting than a more usual statistical approach em-

ployed by Balaji et al. (Balaji et al., 2007).

An incident detection technique using dynamic

time warping is proposed by Hiri-o-Tappa et al. (Hiri-

O-Tappa et al., 2007). Here, the likelihood of con-

gestion is determined using speed data from loop de-

tectors. They authors acknowledge that their strategy

falls short in terms of false alarm and time to detect.

Another approach (Huang et al., 2010) offers a

distributed trafﬁc and congestion detection for au-

tonomous cars. Their approach focuses on wirelessly

connected intelligent vehicles that can measure the

speed of the surrounding trafﬁc and the distance be-

tween the leading and trailing vehicles in order to de-

tect shock waves in the velocity. Their assessment

was based on a highway simulation that was put to

the test under various conditions, such as when an ac-

cident or a road merge were present.

To make short-term predictions of abrupt speed

declines, Labeeuw et al. (Labeeuw et al., 2009) com-

pared their methods to those employing Gaussian Pro-

cesses and decision trees. They used common ma-

chine learning techniques (such as the Support Vector

Machine) as reference. In their evaluation, the deci-

sion tree has the maximum correctness.

3 ORGANIC TRAFFIC CONTROL

Urban road networks typically consist of numerous

signalised intersections which are close to one an-

other. The resulting complexity of dynamic road

trafﬁc patterns, the autonomous behaviour of traf-

ﬁc participants, and the resulting uncertainty offer a

good application for approaches which are based on

the concept of “Organic Computing” (M

uller-Schloer

and Tomforde, 2017). In earlier research, the Ob-

server/Controller architecture was used for the Or-

ganic Trafﬁc Management system (OTC) (Prothmann

et al., 2011b), a self-improving trafﬁc signal control

system.

The key features of this decentralised system in-

clude the capacity to adjust trafﬁc light signalisation

in real-time based observed trafﬁc ﬂows, the estab-

lishment of progressive signal systems (also known as

“green waves”) by contacting adjacent intersections

(Tomforde et al., 2008), and the ability to determine

the most efﬁcient routes to points of interest in the

network and pass on this information to drivers (e.g.

via variable message signs) (Prothmann et al., 2011a).

The number of network-wide stops and journey dura-

tions can be reduced, together with fuel consumption

and pollutant emissions.

Figure 1 outlines the multi-layer OTC architecture

which extends an existing intersection controller (IC)

of a node, the “System under Observation and Con-

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

290

Figure 1: Architecture of the Organic Trafﬁc Control.

trol” (SuOC). Detectors (e.g. loop detectors) measure

the current trafﬁc ﬂows which are handled by Layer 1.

Here, the raw data is preprocessed and combined into

a situation description of intersections which are mon-

itored. The learning classiﬁer system (based on Wil-

son’s XCS (Butz and Wilson, 2002)) then looks for a

matching signalisation plan. In case of a match, sig-

nalisation is adapted accordingly by Layer 0.

The ofﬂine learning feature on Layer 2 is enabled

when an unknown circumstance (where no parame-

ter set is known) arises. The classiﬁer system adds

the new signalisation plan to its knowledge base after

an evolutionary algorithm (EA) generates new signal-

isation plans, analyses them, and sends the best one

back to Layer 1. This layer then responds by exe-

cuting the best ﬁtting signalisation, while Layer 2 ex-

amines alternative approaches. By coordinating their

signalisations, a decentralised collaboration mecha-

nism enables intersection controllers to exchange data

with nearby controllers and construct progressive sig-

nal systems.

3.1 AID Component

The Automatic Incident Detection (AID) component

of OTC is situated in the observer at layer 1. It over-

sees the initial creation of the structural components

required for detection (Fig. 2). A “link” consists of

one or more road segments and their respective detec-

tors. It connects two junctions. A detector pair con-

sists of exactly two loop detectors which are placed

20m apart on the same segment. Additionally, a di-

vided detector pair can monitor longer distances of

roadways more accurately and span branching por-

tions. The part of the network an IC can oversee is

called monitoring zone. Exactly one node is assigned

to each zone. This allows for the intersection-centric

execution of various event detection algorithms with

various parameter settings at each junction. Finally,

an online calibration component can modify param-

eters used by the algorithms according to their false

alarm and detection rates. An incident alarm raised

by the AID component has to be conﬁrmed or marked

as a false alarm by a human expert.

Figure 2: Structural components for the AID.

3.2 Prediction Component

The prediction component on Layer 1 of the rele-

vant node forecasts the trafﬁc ﬂows for each inter-

section’s turning and outgoing sections. The compo-

nent consists of a number of predictors that each in-

dependently anticipate the ﬂow during the next time

step or for a future point in time. Past trafﬁc ﬂows

recorded by detectors and processed and transmitted

by the monitoring component serve as the basis for

these projections. The techniques implemented for

the forecasts range from simple moving average to

more sophisticated algorithms such as the Kalman

ﬁlter (Okutani and Stephanedes, 1984) and Artiﬁ-

cial Neural Networks as suggested in (Sommer et al.,

2013), which may rely on current trafﬁc data as well

as historical data.

All forecasts are then combined into one compre-

hensive forecast based on the previous performance

of each predictor. Thus, the system learns to produce

algorithms with a higher accuracy (lower forecast er-

ror) as well as more inﬂuence on the combined fore-

cast. As seen in Eq. (1), the aggregated forecast is cal-

culated based on the weighted sum of the individual

forecasts F

to F

. Several approaches to deﬁne these

Approaches to Automatic Road Trafﬁc Incident Detection and Incident Forecasting

291

weights as Simple Average (Zhang, 2012), Optimal

Weights (Bates and Granger, 2001), Outperformance

(de Menezes et al., 2000) were implemented.

F =

× F

+ w

× F

+ ··· + w

× F

∑

(1)

These aggregated forecasts, along with the current

scenario from the situation analyser and the perfor-

mance analyser, are given to the situation descriptor,

which then creates a description of the situation. The

controller on Layer 1 is then informed of the situation

and uses the description to choose the ensuing signal-

isation scheme.

4 INCIDENT DETECTION

ALGORITHM: EXTENDED

CALIFORNIA ALGORITHM

The Extended California Algorithm for Arterial Envi-

ronments (ECA) aims to extend the existing Califor-

nia Algorithm (CA) whose decision tree can be seen

in Fig. 3. The nodes serve as a platform for various

experiments in which trafﬁc-related data is compared

to predetermined thresholds. Collecting the upstream

and downstream occupancy data requires two detec-

tor pairs on the same network. The occupancy of the

two detector pair differences are calculated in the al-

gorithm’s initial test. The second test is run when it

reaches a speciﬁc threshold T1.

The relative difference in percentage between the

prior occupancies is determined by this test. The

last test is conducted if the discrepancy exceeds the

T2 threshold. This distinguishes between incidents

and trafﬁc jams which produce comparable trafﬁc

patterns. Here, the relative temporal difference of

the downstream detector’s occupancy values is com-

puted. An incident alarm is raised if the outcome ex-

ceeds the T3 threshold. It is assumed that no incident

is present. Usually the algorithm is run every 20 to 60

seconds.

The CA performs well, is simple to comprehend,

and is expandable. Furthermore, it frequently out-

performs more advanced methods, even today. How-

ever, because it was designed for incident detection on

highways, it has a number of drawbacks, particularly

for arterial routes. This can be further explained and

solved by the ECA. Given that the downstream occu-

pancy is lower than the upstream occupancy, trafﬁc

departing on side roads in the vicinity of two detector

pairs may lead in more false alarms. A downstream

occupancy increase caused by an approaching road

Figure 3: State tree of the California Algorithm.

between two detector pairs consequently lowers the

detection rate.

These characteristics, along with signalisation

phase impacts, cause occupant values to ﬂuctuate

rapidly. Due to the brevity of the false alerts and

in particular during periods of low trafﬁc, the occur-

rences reported by the CA are very certainly false

alarms. Finally, because it was created for incident

detection on links, the CA is unable to identify oc-

currences on junctions. These short-comings are ad-

dressed by the ECA:

• Side road check: Consider incoming and outgoing

side roads.

• Incident persistence check: Buffer incident

alarms and report only those which occur again

in the next check.

• Trafﬁc condition check: Introduction of a silent

mode during low trafﬁc situations. Trafﬁc data

is still recorded, but no incident detection alarms

will be forwarded.

• Junction monitoring: Extension of the CA to de-

tect incidents on junctions.

As seen in Fig. 4, the ECA extends the decision

tree of the CA with further states. The ﬁrst state ex-

amines the sudden drop in trafﬁc ﬂow, which is usu-

ally indicative of an accident. The second state is en-

tered if the drop exceeds or is equal to a threshold.

Based on the occupancy difference, this state seeks to

prevent false alarms caused by low trafﬁc density. The

ﬁnal state checks for a persistent drop in occupancy

during an incident. In case of an positive response, an

incident alert is set off. To achieve higher accuracy,

this ﬁnal test may be repeated for greater assurance at

the expense of additional detection time.

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

292

Figure 4: State tree of the ECA.

5 AUTOMATIC INCIDENT

FORECASTING: SUPPORT

VECTOR MACHINE

The limitation of incident detection is that it can only

identify an incident or a congestion if they are already

there. Because congestion is unwanted, it would be

desirable to foresee impending interruptions in the

trafﬁc ﬂow. To this end, an algorithm for the forecast-

ing of trafﬁc jams was created. It relies on short-term

trafﬁc ﬂow projections and the dynamic estimation of

road capacities. The theory behind this method is that

congestions can be anticipated if the current trafﬁc

ﬂow exceeds the capacity of the road. In case of a

bottleneck, it is assumed that the incident has ended

once the ﬂow increases again. This is depicted in in

Fig. 5.

Figure 5: Concept of the incident forecast algorithm.

5.1 Dynamic Capacity Calculation

The dynamic calculation of the capacity of a section

follows the formula provided by the U.S. Department

of Transportation (FHWA (Federal Highway Admin-

istration), 2010) which takes into consideration e.g.

the current speed limit, the number of lanes and the

percentual ratio of heavy vehicles. This formula dis-

tinguishes between highway and urban roads, based

on the speed limit. Equation (2) depicts the formula

for the dynamic capacity calculation in urban areas

where S

is the base ﬂow rate, N is the number of

lanes, f

is a factor for the lane width, F

is the fac-

tor for heavy vehicles, f

is an adjustment factor for

parking activities, f

is an adjustment factor for the

area type and PHF is the peak hour factor. Thus, a

closed lane or a higher ratio of heavy vehicles will

lead to a lower capacity.

cap = S

∗ N ∗ f

∗ f

∗ PHF (2)

5.2 Incident Forecast Algorithm

The developed incident forecast algorithm follows the

idea of state-based algorithms. Its decision tree is de-

picted in Fig. 6. State 0 stands for “incident free”,

state 1 for “tentative incident”, state 2 for “incident

conﬁrmed” and state 3 for “incident continuing”.

As before, C

holds the occupancy value of a de-

tector, f low

t+∆t

is the trafﬁc ﬂow forecast at ∆t steps

into the future and i is the time until the next algo-

rithm execution. In an “incident free” state, the fore-

casted ﬂow is compared to the dynamic capacity: If

the forecast is higher, the current state is switched to

“tentative incident”. In the next iteration, the algo-

rithm evaluates the left side of the decision tree, run-

ning the same test again; in case of positive feedback,

the state is changed to “incident conﬁrmed”. Finally,

the lower left state is executed, which checks if the

forecasts are again lower than the dynamic capacity

and subsequently raise once more. If this is true, we

assume that the incident is dissolving. Otherwise, the

incident is continuing.

Figure 6: Decision tree of the AID forecast algorithm.

Approaches to Automatic Road Trafﬁc Incident Detection and Incident Forecasting

293

6 EVALUATION

The evaluation was conducted using the OTC soft-

ware with its AID capabilities together with “Aimsun

Next 22.0” (Aimsun, 2022), a professional trafﬁc sim-

ulator employed in the ﬁeld of trafﬁc engineering. It

was used to simulate parts of a real-world network.

6.1 Network Model

Figure 7 depicts the simulated roads of an urban net-

work in Hamburg, Germany. It consists of 10 junc-

tions which are each equipped with a ﬁxed-time TLC.

The simulated trafﬁc demand is based on trafﬁc states

which describe the hourly ﬂow per road. Here, a

work-day morning from 6:00 to 12:00 is modelled,

to emulate a rush-hour with a varying and increased

trafﬁc demand of 62840 vehicles per hour. The inci-

dents simulated for the evaluation are located on the

section from junction K5 to junction K10 (see Fig. 8).

As K10 is not controlled by a TLC, the monitoring

zone according to Section 3.1 encompasses the links

with all sections from K5 westward to K6.

Figure 7: The simulated part of the road network in Billst-

edt, an urban district of Hamburg, Germany.

6.2 Experiments

Metrics such as the detection rate (DR), the false

alarm rate (FAR) and the average time to detection

(ATTD) were used for the evaluation:

DR =

Number of detected incidents

Real number of incidents

(3)

FAR =

Number of false alarms

Number of reported incidents

(4)

AT T D =

∑

i=1

detection

−t

occurence

(5)

The ATTD is the average over the sum of the

differences between the detection time and the real

starting time of the incident. FAR and DR are met-

rics which measure the effectiveness of an algorithm,

whereas ATTD reﬂects the efﬁciency of an algorithm.

To emulate typical situations, these metrics are ap-

plied in the context of two experimental scenarios:

1. “Section closure”: Sudden blockage of a whole

section, by closing all lanes for 20m at some loca-

tion within the section.

2. “Partial lane closure”: Sudden blockage of one or

more lanes of a section (but not all), again not for

the whole section length, but for 20m

Every hour (starting with 7:00–7:15), the inci-

dents are each simulated at the 4 different loca-

tions depicted in Fig. 8. With varying random seeds

these simulations are each repeated 5 times to obtain

pseudo-randomised replications. Finally, this is done

for each of the algorithms.

6.3 Results

The overall goal is to increase FAR and ATTD while

increasing DR. Unfortunately, these goals interfere

with each other. A higher DR comes a the price of

a higher FAR while a reduced DR leads to a lower

FAR. The tuning of the parameters (time between

algorithm executions, speciﬁc conﬁguration parame-

ters, etc.) of each incident detection algorithm is im-

portant to achieve a good performance.

For the ﬁrst scenario, Table 1 outlines some ﬁnd-

ings. The California 7 algorithm has difﬁculties to

detect the fairly pronounced effects of the section clo-

sures of this setting and reports false alarms (mainly

due to false locations). California 8 fares better due to

its analysis of compression waves. The detection for

both is quick. The ECA is slower, but shows no prob-

lems concerning correct detections or false alarms.

This is valid for the faster forecast algorithm.

Table 1: Results of the ﬁrst scenario “section closures” as

means over 5 replications.

DR FAR ATTD [s]

California 7 50% 70% 60

California 8 100% 33% 60

ECA 100% 0% 120

Forecast 100% 0% 72

In the second scenario (Table 2), the overall de-

tection declines. This is somewhat expected as the

trafﬁc ﬂow in sections with single-lane incidents is

not substantially different from free-ﬂowing condi-

tions. Still, all algorithms detect some incidents cor-

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

294

Figure 8: Part of the network in Fig. 7 which features the simulated detector pairs A,B and C,D as well as E,F. The locations

of the simulated incidents are labelled a, b, c and d.

rectly, although with low detection rates and in the

case of the California algorithms, which also show

high false positive rates. Again, the forecast algorithm

performs better in terms of detection, although it takes

the longest.

Table 2: Results of the ﬁrst scenario “partial lane closures”

as means over 5 replications.

DR FAR ATTD [s]

California 7 20% 40% 192

California 8 25% 80% 135

ECA 20% 0% 210

Forecast 75% 0% 242

7 CONCLUSION

In this paper, two algorithms for the detection and pre-

diction of incidents were presented. The Enhanced

California algorithm was evaluated for different sce-

narios and performed better when compared with the

other California Algorithm 7 and 8 algorithms. For

the less pronounced congestions of partial lane clo-

sure, the capacity-based forecast algorithm outper-

forms the others, as they show insufﬁcient detection

rates. Future work could include extended test sce-

narios with additional types of incidents and trafﬁc

demands to consolidate these ﬁndings.

Further research aims to use the incident informa-

tion in the routing component of the OTC. It would

then be able to issue optimised routing recommen-

dations for the drivers to prevent congestions and

to accelerate the breakup of the congestion. An

incident-aware adaptation of the trafﬁc light signalisa-

tion could also beneﬁt the overall goal of a proactive

prevention of congestion.

REFERENCES

A.I., G. and Hall, F. (1989). Distinguishing between inci-

dent congestion and recurrent congestion: A proposed

logic. Transportation Research Record 1232, pages

1–8.

Aimsun (2022). Aimsun Next 22 User’s Manual. Barcelona,

Spain, Aimsun Next 22.0.1 edition.

Balaji, P. G., Sachdeva, G., Srinivasan, D., and Tham, C.-K.

(2007). Multi-agent system based urban trafﬁc man-

agement. IEEE Congress on Evolutionary Computa-

tion, pages 1740 – 1747.

Bates, J. M. and Granger, C. W. J. (2001). The Combination

of Forecasts, page 391–410. Cambridge University

Press, USA.

Butz, M. and Wilson, S. (2002). An algorithmic description

of XCS. Soft Computing - A Fusion of Foundations,

Methodologies and Applications, 6:144 – 153.

de Menezes, L. M., Bunn, D. W., and Taylor, J. W. (2000).

Review of guidelines for the use of combined fore-

casts. European Journal of Operational Research,

120(1):190–204.

ElSahly, O. and Abdelfatah, A. (2022). A systematic review

of trafﬁc incident detection algorithms. Sustainability,

14(22).

FHWA (Federal Highway Administration) (2010). High-

way performance monitoring system ﬁeld manual.

Guin, A. (2004). An Incident Detection Algorithm based on

a Discrete State Propagation Model of Trafﬁc Flow.

PhD thesis, Georgia Institute of Technology.

Hiri-O-Tappa, K., Narupiti, S., Pan-Ngum, S., and Pattara-

Atikom, W. (2007). Development of real-time short-

term trafﬁc congestion prediction method. Journal of

Society for Transportation and Trafﬁc Studies, 2.

Huang, D., Shere, S., and Ahn, S. (2010). Dynamic high-

way congestion detection and prediction based on

shock waves. In Proc. of Inte. Works. on Veh. Inter-

NETworking, VANET ’10, pages 11–20, New York,

NY, USA. ACM.

Kurihara, S. (2013). Trafﬁc-congestion forecasting algo-

rithm based on pheromone communication model. In

Approaches to Automatic Road Trafﬁc Incident Detection and Incident Forecasting

295

Barbosa, H. J., editor, Ant Colony Optimization, chap-

ter 7. IntechOpen, Rijeka.

Labeeuw, W., Driessens, K., Weyns, D., Holvoet, T., and

Deconinck, G. (2009). Prediction of Congested Trafﬁc

on the Critical Density Point Using Machine Learning

and Decentralised Collaborating Cameras. In Proc. of

EPIA’09, pages 15–26. Universidade de Aveiro.

Marchesini, P. and Weijermars, W. A. M. (2010). The re-

lationship between road safety and congestion on mo-

torways. SWOV Institute for Road Safety Research

Leidschendam.

Masters, P., Lam, J., and Wong, K. (1991). Incident de-

tection algorithms for compass - an advanced trafﬁc

management system. In Vehicle Navigation and In-

formation Systems Conference, 1991, volume 2, pages

295–310.

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

puting - Technical Systems for Survival in the Real

World. Birkh

auser.

Okutani, I. and Stephanedes, Y. J. (1984). Dynamic pre-

diction of trafﬁc volume through kalman ﬁltering the-

ory. Transportation Research Part B: Methodological,

18(1):1–11.

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

cident detection algorithms & their deployment: What

works and what doesn’t. Technical report, Fall River,

MA: New England Transp. Consortium.

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

detection algorithms based on decision trees with

states. Transportation Research Record, No. 682,

pages 30–37.

Prothmann, H., Schmeck, H., Tomforde, S., Lyda, J.,

ahner, J., M

uller-Schloer, C., and Branke, J. (2011a).

Decentralised route guidance in organic trafﬁc con-

trol. In Proc. of SASO 2011, pages 219–220. IEEE

Computer Society.

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

ahner, J.,

uller-Schloer, C., and Schmeck, H. (2011b). Or-

ganic trafﬁc control. In Organic Computing – A

Paradigm Shift for Complex Systems, chapter 5.1,

pages 431–446. Birkh

auser Verlag.

Rao, A. M. and Rao, K. (2012). Measuring urban trafﬁc

congestion – a review. Int. J. for Trafﬁc and Transport

Engineering, 2(4):286–305.

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

ahner, J. (2013). Using

a neural network for forecasting in an organic trafﬁc

control management system. In Proc. of ESOS ’13.

Tomforde, S. (2012). Runtime adaptation of tech-

nical systems: An architectural framework for

self-conﬁguration and self-improvement at runtime.

udwestdeutscher Verlag f

ur Hochschulschriften.

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 Proc. of SASO 2008, pages

413–422. IEEE Computer Society.

Zhang, Y. (2012). How to provide accurate and robust traf-

ﬁc forecasts practically? In Abdel-Rahim, A., edi-

tor, Intelligent Transportation Systems, chapter 8. In-

techOpen, Rijeka.

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

296