Wprowadzanie ulic szkolnych: perspektywa uczenia maszynowego

Tomasz Dzieduszyński, Olga Czeranowska-Panufnik

doi:10.37190/arc250113

Streszczenie

Tematem artykułu są ulice szkolne, a zwłaszcza proces ich tworzenia z perspektywy uczenia maszynowego. Autorzy przedstawili w nim możliwości zastosowania generatywnych sieci przeciwstawnych typu GAN we wdrażaniu koncepcji zamknięcia ulic szkolnych, której celem jest poprawa bezpieczeństwa i zmniejszenie natężenia ruchu wokół szkół. Przeanalizowali programy ulic szkolnych na całym świecie, identyfikując powtarzające się wyzwania i proponując rozwiązania. Na podstawie analizy 51 skutecznych wdrożeń oraz nowej metody ekstrakcji cech miejskich opracowali model uczenia maszynowego, który wspomaga wybór potencjalnych lokalizacji ulic szkolnych. Tak przygotowane narzędzie ma służyć usprawnieniu procesu wyboru i zwiększeniu efektywności projektu przez lepsze dopasowanie do lokalnego kontekstu. Pomimo pewnych ograniczeń, takich jak trudność w odwzorowaniu wszystkich kontekstów przestrzennych, system dostarcza wartościowych wniosków dotyczących regulacji ruchu miejskiego. Badanie wypełnia lukę w literaturze naukowej i oferuje podejście oparte na danych do projektowania bezpieczniejszych przestrzeni miejskich.

Słowa kluczowe:
ulice szkolne
uczenie maszynowe
smart city
generatywne sieci przeciwstawne
projektowanie architektoniczne wspomagane komputerowo

Podgląd pełnego artykułu jest możliwy wyłącznie na większych ekranach.

2025

1(81)

Tomasz Dzieduszyński*, Olga Czeranowska-Panufnik**

School streets implementation: A machine learning perspective

DOI: 10.37190/arc250113

Published in open access. CC BY NC ND license

Abstract

This study explores the application of generative adversarial neural networks in the implementation of school street closures, a concept aimed at

improving safety and reducing trac around schools. The research analyses various school street closure programs worldwide and identies common

challenges. Based on an analysis of 51 successful school street closure programs and a proposed urban feature extraction method, a supervised machine

learning model was developed to facilitate the selection of the potential school street candidates. The developed tool aims to streamline the selection

process and make the design more eective and context-sensitive. Despite certain limitations, such as the inability to represent all spatial contexts

and small-scale urban details, the system can propose a meaningful denition of a school street closure zone. This research contributes to the limited

academic literature on school streets and play streets, providing a new perspective on urban trac regulation.

Key words: school streets, machine learning, smart city, generative adversarial networks, computer aided architectural design

Introduction

School Street is a road outside a school with a tempo-

rary restriction on motorized trac at school drop-o and

pick up times (School Street Initiative 2019). The idea to

close down the trac for the sake of children is already

over one hundred years old and can be traced back to the

idea of “play streets” implemented in Cincinnati as early

as in 1920 (Reeves 1931). However, the modern under-

standing of the concept of school streets is relatively new

in urban studies, as the rst contemporary school streets in

Europe were implemented in the late 1980s in Bolzano, It-

aly (Brand, Böhler and Rupprecht 2021; Pressl 2011). The

idea is insuciently explored in peer-reviewed literature

(Davis 2020). There is a gap between the need for their

implementation and the theoretical foundations as well

as available tools, especially in Poland, where the school

street closure programs are just in their infancy.

This study aims to analyse contemporary worldwide

examples of school streets to identify issues in the stud-

ied pilot programs and develop a machine learning tool for

selection and analysis of school streets candidates based

on their successful worldwide implementations. The study

is divided into three main parts: literature review supple-

mented by qualitative analysis of school street closure pro-

grams, denition of a training dataset based on the selected

positive examples paired with the corresponding urban fea-

tures, and development of a supervised machine learning

algorithm based on generative adversarial network (GAN)

architecture (Goodfellow et al. 2014; Isola et al. 2017).

State of the art, school street programs review

The academic sources on both school streets and play

streets are limited (Davis 2020; Bridges et al. 2020). A sup-

plementary Web of Science query from April 2024 for

the tags “school streets” or “school street” returned only

10 re levant peer reviewed publications (Clarivate 2024).

The sources on the subject consist mostly of grey litera-

ture pilot program reports, non-peer-reviewed studies,

sur vey results, local press reports and practical implemen-

tation toolkits. The present study analyses a selection of

school

street closure programs from the years 2016 to 2022

* ORCID: 0000-0003-2571-5128. Faculty of Architecture, War-

saw University of Technology, Poland, e-mail: tomasz.dzieduszynski@

pw.edu.pl

** ORCID:

0000-0002-8210-5265. Faculty of Architecture, War -

saw

University of Technology, Poland.

140 Tomasz Dzieduszyński, Olga Czeranowska-Panufnik

localized mostly in Great Britain, the current world lead-

er in the number of school street closure zone implemen-

tations (Clarke 2022). The conducted analysis included

London (Noble et al. 2021), Hackney (Linton et al. 2021),

Edinburgh (Lawrence, Murrell 2016; City of Edinburgh

Council 2020), Southampton (My Journey 2021), East Lo-

thian (Stubbs 2016; Ritchie 2016), Solihull (Keaney, Tov-

ey 2018; Solihull Metropolitan Borough Council 2021),

Perth and Kinross council area (Perth and Kinross Council

2022), but also Victoria, Canada (880 Cities 2019; City of

Victoria 2022), Warsaw, Poland (Mikołajczyk 2021; Urząd

m.st. Warszawy 2022) and Wrocław, Poland (Grabowska,

Szmigiel-Franz 2020).

In the analysed grey literature the implementation of

school streets was mainly motivated by safety concerns

(Fig. 1), but the premises encompassed a variety of factors.

In Southampton, the program aimed to enhance safety and

air quality. It promoted outdoor exercise and encouraged

alternative transportation like bikes and scooters, while

also reducing trac congestion (My Journey 2021). East

Lothian’s program aimed to increase road safety by reduc-

ing trac, congestion and air pollution near school gates.

It also encouraged active travel among students and their

families and fostered an environment for social interaction

(Stubbs 2016). Solihull’s initiative was introduced to ad-

dress increased trac during school drop-o and pick-up

times, which often led to potentially dangerous situations

caused by the drivers (Keaney, Tovey 2018). In London

the school street initiative gained popularity thanks to the

“Streetspace for London” program launched as an emer-

gency response to the COVID-19 pandemic. The program

aimed to make social distancing easier while traveling in

London, encourage people to walk, cycle and scoot, pre-

vent an increase in car use enabling emergency services

and essential vehicle journeys and to enhance the air quali-

ty. As a result of the program, over 300 school streets were

implemented in London (Noble et al. 2021). Generally,

the COVID-19 pandemic vastly accelerated the develop-

ment of school street programs more than doubling the

worldwide number of cities participating in the initiative

(Clarke 2022).

In the analysed sources, the selection and design pro-

cess of the school street closure zones was based on several

urban criteria, mostly focused on the road network sur-

rounding the candidate school. In East Lothian, factors such

as road type, existing speed restrictions or trac regulation

orders were considered, as well as the recorded accident

history and availability of larger municipal parking. Addi-

tionally, the preferences and habits of the students and their

parents were recognized through the analysis of the number

of students dependent on road transport (Stubbs 2016). In

Solihull, UK, the selection was based on the type of road,

the absence of more intense throughfares and the feasibil-

ity of trac reorganization. The supplementary contexts

consisted of the neighbourhood structure, the percentage of

students living within the walking distance

from the school

and the willingness of the school and the local community

to cooperate (Keaney, Tovey 2018). In Wrocław, Poland, the

criteria focused on the possibility of practical trac orga-

nization to bypass the closed section, the registered reports

of dangerous situations or increased trac during school

hours and previous participation of a candidate school in the

programs implemented by the Sustainable Mobility Oce

(Biuro Zrównoważonej Mobilności). The character of the

neighbourhood was also considered, including the distribu-

tion of students living within the 15-minute radius from the

school and the accessibility to public transport (Grabowska,

Szmigiel-Franz 2020).

Overall, the analysed school street pilot programs were

mostly successful. In the reviewed reports 51 examples of

implementations were recognized as unambiguously posi-

tive, and only 5 school street pilots failed and had to be ter-

minated (Fig. 2). However, even the successful programs

encountered some problems. In certain cases, the program

could be implemented more strictly and the trac regula-

tion orders could be better executed. Some drivers were still

speeding or passing through the street during a closure. To

some extent this was often caused by confusion of drivers

not aware of the trac regulation order in place. The imple-

mented regulations were often chaotic. It was not obvious

when the restrictions were active, as they were not consis-

tent across dierent months or days of the week. It was also

not clear who was exempt from the restrictions (Keaney,

Tovey 2018). To make the situation clearer, in some cities

parents or the school were responsible for setting up a scis-

sor gate. The school street could also be monitored by a city

Fig. 1. The geometric intuition behind the school street closure zones (red) near the school access points (blue).

Increase in the number of drop-off nodes (green) reduces the related traffic problems and enhances safety

(elaborated by T. Dzieduszyński, O. Czeranowska-Panufnik)

Il. 1. Geometryczna intuicja stojąca za strefami zamknięcia ulic szkolnych (czerwone linie) w pobliżu punktów dostępu do szkoły (niebieskie punkty).

Zwiększenie liczby punktów wysiadania z pojazdu (zielone punkty) zmniejsza związane z tym problemy komunikacyjne i zwiększa bezpieczeństwo

(oprac. T. Dzieduszyński, O. Czeranowska-Panufnik)

School streets implementation: A machine learning perspective 141

guard (Keaney, Tovey 2018; Grabowska, Szmigiel-Franz

2020). In many cases, the implementation of school streets

impacted many stakeholders of dierent needs and interests

(Grabowska, Szmigiel-Franz 2020). The programs required

extensive informational campaigns and personal consulta-

tions to be supplemented. Moreover, the closures often bur-

dened the inhabitants of school streets. While local residents

typically received a certain number of passes, some com-

plained about the inability to receive guests during the clo-

sures (Keaney, Tovey 2018). Additionally, these programs

heavily burdened parents, necessitating sudden changes of

routines. Some parents even questioned the necessity of the

programs and believed that the reasons for their implemen-

tation were non-existent. In one of the reviewed cases, the

way to school became perceivably more dangerous after

the implementation of a school street closure zone, since

children had to cross a large road (Keaney, Tovey 2018).

A few critics of the idea raised a claim that the problems

of chaotic parking, dangerous trac and air pollution were

not solved by the closure programs, but were just moved

to another place or caused a “wave eect” (Grabowska,

Szmigiel-Franz 2020; Lawrence, Murrell 2016). While this

could be true for incorrectly designed trac regulation or-

ders, a school street closure zone with numerous, well-de-

signed drop-o nodes would have a tendency to better dis-

tribute these issues across a broader area, thus decreasing

their severity (Fig. 1). To a large extent the mentioned prob-

lems were caused by the insucient consideration of the

contexts and available data, as well as too hermetic social

consultations. A wider implementation of data-driven tools

(such as the one proposed in this study) could potentially al-

leviate the negative eects of school street implementations

related to the lacking recognition of local urban contexts.

Machine learning perspective

Following a comprehensive review of school street

closure programs implemented between 2016 and 2022,

a prototype machine leaning context-analysis algorithm

was developed in 2023 and 2024. This algorithm was

trained on the selection of 51 successful school street pi-

lot programs identied through grey literature analysis.

During inference, the prototype uses the coordinates of

a candidate school. It then extracts the most relevant spatial

contexts from OpenStreetMap (OpenStreetMap Contribu-

tors 2024) and translates these contexts into the proposed

closure zone, together with the proposed drop-o nodes

and school access points. In its current iteration, the tool

functions as a preliminary analysis instrument, focusing

solely on readily accessible open-source urban contexts.

This approach avoids the need for additional viability and

social analysis. These contexts are compiled in the form of

three voxel maps representing the urban tissue (buildings,

car parks, bridges), extents of the analysed school (school

area) and the communication network (roads, paths, tun-

nels and waterways). These maps can be represented as

a 512 × 512 RGB image, where each of the colour maps

represents one of the context groups and the intensity of

the pixel’s colour corresponds to the height of the urban

tissue and the intensity of the transit mode on the com-

munication network map (Fig. 3). One pixel represents an

area of 1.5 × 1.5 m, the extent of the context map corre-

sponds to the pedestrian shed dened by a 5-minute walk

radius, approximately 400 m (Perry 1929).

The algorithm is based on a generative adversarial neural

network (GAN) (Goodfellow et al. 2014), a type of neural

network used for generating data similar to a given dataset.

Its structure was inspired by the pix2pix architecture (Iso-

la et al. 2017). From the conducted school street programs

review, 51 the most successful examples of closure zones

were selected and RGB target maps were manually labelled.

The red channel on the target map represents the closure

zone, the green represents the drop-o node and blue rep-

resents the school access point. Coordinates of the selected

schools were compiled in an XML spreadsheet. A custom

script was used to automatically download the urban fea-

ture maps from the Mapbox Static Image Service (Mapbox

2023) through API and pair them with the hand-labelled

target maps. The training dataset was augmented through

random cropping, random jitter and random mirroring to

reduce the risk of overtting. U-Net (Ronneberger, Fischer

and Brox 2015) was used to generate the proposed school

street closure zone based on the given feature map. A patch

discriminator was applied to try to distinguish the generated

maps from the original target maps. The L1 and GAN loss

were calculated based on the algorithm’s performance and

Fig. 2. The ratio of successful

school street implementations

to failed ones based on the grey

literature analysis

(elaborated by T. Dzieduszyński,

O. Czeranowska-Panufnik)

Il. 2. Stosunek udanych wdrożeń

programów zamknięcia ulic

szkolnych do nieudanych,

na podstawie analizy

literatury szarej

(oprac. T. Dzieduszyński,

O. Czeranowska-Panufnik)

142 Tomasz Dzieduszyński, Olga Czeranowska-Panufnik

fully generated meaningful school street closure zones in

90% of validation cases (Fig. 5). The inference time on

a mid-range consumer laptop ts in the range of 1–3 s,

which allows for automated batch-processing of large sets

of candidate schools.

Case study

In this section, we provide detailed examples of the

al gorithm applied to real-world scenarios, specically in

Warsaw’s Elementary School No. 195 and No. 175. By ad-

justing road class parameters and rerunning the algorithm,

we explored dierent closure scenarios for these schools.

back-propagation was used to train both of the networks in

an ensemble (Fig. 4). Hyperparameters were tuned using

the random search method. The lambda hyperparameter

was tuned during training, increasing the weight of the L1

loss function in the later phases of training. The algorithm

was trained for 31 hours on Nvidia GF RTX 2080 Super.

During inference only the generator is used.

The developed algorithm was tested on the selection

of candidate schools for the Warsaw “Droga na szóstkę”

school street viability evaluation program (LPW 2020), as

well as four negatively assessed schools described in the

analysed pilot reviews. These schools were not included

in the training data. The trained neural network success-

Fig. 3. The context map represented as an RGB image. The green channel represents the height of the urban tissue,

blue channel represents the extents of the analysed school land, the red channel represents the intensity of the communication network.

The map is extracted from OpenStreetMap (OpenStreetMap Contributors 2024) through Mapbox Static Images Service (Mapbox 2023)

(elaborated by T. Dzieduszyński, O. Czeranowska-Panufnik)

Il. 3. Mapa kontekstów przedstawiona jako obraz RGB. Kanał zielony reprezentuje wysokość tkanki urbanistycznej,

kanał niebieski przedstawia zasięg terenu analizowanej szkoły, a kanał czerwony intensywność sieci komunikacyjnej.

Mapa została pozyskana z OpenStreetMap (OpenStreetMap Contributors) za pośrednictwem usługi Mapbox Static Images (Mapbox 2023)

(oprac. T. Dzieduszyński, O. Czeranowska-Panufnik)

Fig. 4. The training diagram of the utilized generative adversarial neural network (elaborated by T. Dzieduszyński, O. Czeranowska-Panufnik)

Il. 4. Diagram przedstawiający proces treningu zastosowanej generatywnej sieci przeciwstawnej GAN

(oprac. T. Dzieduszyński, O. Czeranowska-Panufnik)

School streets implementation: A machine learning perspective 143

The main purpose of the developed tool is the facilita-

tion of the school street selection process. The algorithm

receives the coordinates of the queried school, extracts the

urban contexts and in a few seconds, it outputs the pro-

posed school street closure zone. The generated proposal

can be used to assess the sensibility of the closure zone im-

plementation and can be later used as a baseline during the

later stages of evaluation and design process. The second

use case is the urban redesign analysis. The input feature

maps do not have to correspond to the current urban con-

texts, but can represent the planned urban interventions.

These interventions could relate to road classes, number

of car parks, redesign of the urban tissue, changes in pe-

destrian paths network or change in development intensity.

The varying input feature maps result in dierent target

maps generated by the algorithm and can help urban de-

signers better understand the relationships between school

streets and their surroundings. This use case was tested

and demonstrated on the example of changing road classes

surrounding the Warsaw Elementary School No. 195 and

No. 175. Lowering the class of the road results in the ex-

tension of the possible closure zone, while increasing the

class leads to its reduction. To compensate for the shrink-

ing of the school street, the algorithm generates additional

peripheral drop-o nodes, thereby enhancing the ecien-

cy of the closure program (Fig. 6).

The results demonstrate the adaptability of the tool in

urban design. For example, lowering the road class sur-

rounding Elementary School No. 195 signicantly expand-

ed the closure zone, while increasing the road class around

No. 175 led to more distributed peripheral drop-o nodes.

These applications demonstrate how changes in urban fea-

tures aect the overall design, providing a exible solution

for urban planners.

Discussion

As demonstrated, the system can propose meaningful

school closure zones. This can be benecial during the

selection process of school street candidates and can be

utilized during urban redesign analysis. Although the re-

sults are promising, the developed tool is currently only

a prototype and it has many limitations, which reduce its

practical applicability in urban planning. One of the key

limitations is that the tool does not fully reect the com-

plexity of the school street selection process. It does not

account for important social contexts, such as stakehold-

er dynamics, local policies or preferences of students and

Fig. 5. School street closure zones generated by the trained algorithm for the three Warsaw candidates for the school street closure program (LPW 2020):

a) Elementary School No. 175, b) Elementary School No. 157, c) Elementary School No. 373.

Red channel of the target space represents the closure zone, green the drop-off nodes and blue the school access point

(elaborated by T. Dzieduszyński, O. Czeranowska-Panufnik)

Il. 5. Strefy zamknięcia ulic szkolnych wygenerowane przez wytrenowany algorytm dla trzech warszawskich kandydatów

do programu ulic szkolnych (LPW 2020): a) Szkoła Podstawowa nr 175, b) Szkoła Podstawowa nr 157, c) Szkoła Podstawowa nr 373.

Kanał czerwony przestrzeni celu przedstawia strefę zamknięcia, zielony – punkty wysiadania z pojazdu, a niebieski punkt dostępu do szkoły

(oprac. T. Dzieduszyński, O. Czeranowska-Panufnik)

a b c

144 Tomasz Dzieduszyński, Olga Czeranowska-Panufnik

their families. These factors play a signicant role in the

successful implementation of school street closures, and

their absence in the model reduces the sensitivity of the

tool to such nuances. It is important to note that the tool

was not developed to replace the complex urban analyses

required for these projects but to complement them by pro-

viding an additional layer of analysis that makes the design

process more context-sensitive.

The proposals generated by the neural network were

some times hard to interpret. By machine learning standards,

the training sample was very small, consisting of only

successful training examples represented in the reviewed

literature. This problem was partially mitigated by data

augmentation, but overall, the small training sample signi-

cantly reduced the algorithm’s generalizability, as it could

not encounter many context scenarios during the training.

Such case can be visualized on the example of the Bonaly

Primary School in Edinburgh, which closely neighbours

the Edinburgh highway bypass. Since this specic situation

was not represented in the training set, the algorithm de-

cided to generate the drop-o points in the middle of the

highway (Fig. 7).

Fig. 6. Utilization of the trained neural network in urban redesign analysis demonstrated on:

a) Warsaw Elementary School No. 195, b) Warsaw Elementary School No. 175.

Changes in road classes affect the proposed school street closure zones (elaborated by T. Dzieduszyński, O. Czeranowska-Panufnik)

Il. 6. Wykorzystanie wytrenowanej sieci neuronowej w analizie przekształceń urbanistycznych pokazane na przykładach:

a) warszawskiej Szkoły Podstawowej nr 195, b) warszawskiej Szkoły Podstawowej nr 175.

Zmiany w klasach dróg wpływają na proponowane strefy zamknięcia ulic szkolnych (oprac. T. Dzieduszyński, O. Czeranowska-Panufnik)

Fig. 7. An example of an incorrect school street closure program proposed by the algorithm for the Bonaly Primary School

resulting from the small training sample. The algorithm generated the drop-off nodes in the middle of a highway

(elaborated by T. Dzieduszyński, O. Czeranowska-Panufnik)

Il. 7. Przykład nieprawidłowego programu zamknięcia ulicy szkolnej zaproponowanego przez algorytm dla Szkoły Podstawowej Bonaly,

wynikający z niewielkiej próbki treningowej. Algorytm wygenerował punkty wysiadania z pojazdu na środku autostrady

(oprac. T. Dzieduszyński, O. Czeranowska-Panufnik)

School streets implementation: A machine learning perspective 145

References

880 Cities. School Streets Guidebook. City of Victoria, Capital Regional

District of British Columbia, 2019.

Brand, Lasse, Susanne Böhler, and Siegfried Rupprecht. Topic Guide:

Sustainable Urban Mobility Planning in Smaller Cities and Towns.

Rupprecht Consult, 2021.

Bridges, Christina N., Tyler M. Prochnow, Emily C. Wilkins, Ke shia

M. Pollack Porter, and M. Renée Umstattd Meyer. “Examin-

ing the Implementation of Play Streets: A Systematic Review of

the Grey Literature.” Journal of Public Health Management and

Practice 26, no. 3 (2020): E1–E10. https://doi.org/10.1097/PHH.

0000000000001015.

City of Edinburgh Council. “Streets Ahead, Road Safety in Edinburgh.

School Streets.” Accessed April 10, 2020, at https://www.streetsa-

headedinburgh.org.uk/school-streets-1/school-streets.

City of Victoria. “School Streets. The School Streets program creates

safer school zones and encourages active transportation.” City of

Vic toria. 2022. Accessed April 10, 2022, at https://www.victoria. ca/ -

EN/main/residents/streets-transportation/walk-roll-transit/school-

programs/school-streets.html.

Clarivate. Web of Science. n.d. Accessed April 10, 2024, at https://www.

webofscience.com/wos/woscc/summary/6d7384cd-379f-477d-

9581-eb2df65998e9-e194a1f6/relevance/1.

Clarke, Richard. School streets: Putting children and the planet rst.

Child Health Initiative’s Advocacy Hub, FIA Foundation, 2022.

Davis, Adrian. School Street Closures and Trac Displacement Project:

A Literature Review with semi-structured interviews. Transport Re-

search Institute, Edinburgh Napier University, 2020.

Goodfellow, Ian J., Jean Pouget-Abadie, Mehdi Mirza, et al. “Genera-

tive Adversarial Nets.” In Advances in Neural Information Process-

ing Systems (NIPS), edited by Zoubin Ghahramani, Max Welling,

Corinna Cortes, Neil D. Lawrence and Kilian Q. Weinberger. Curran

Associates, 2014.

Grabowska, Marta, and Anna Szmigiel-Franz. Szkolna ulica: Raport

z pilotażu. Biuro Zrównoważonej Mobilności Urzędu Miejskiego

Wrocławia, 2020.

Isola, Phillip, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. “Image-to-

-Image Translation with Conditional Adversarial Networks.” In IEEE

Conference on Computer Vision and Pattern Recognition (CVPR),

2017, 5967–5976.

Keaney, David, and Paul Tovey. School Streets Pilot Project – Six Month

Update and Review. Report, Solihull Metropolitan Borough Coun-

cil, 2018.

Lawrence, Paul, and Steven Murrell. School Streets Pilot Project Eva

luation. Transport and Environment Committee, The City of Edin-

burgh Council, 2016.

Linton, Tyler, Doolin O’Reilly, Dominique Humbert, and Caroline Bur-

well. School Streets Timed Trac Restrictions Toolkig for Profes-

sionals. Hackney School Streets, 2021.

LPW. “Droga na szóstkę: Ocena bezpieczeństwa ruchu drogowego wo-

kół wybranych szkół podstawowych: Raport końcowy, Szkoła pod-

stawowa nr 41; 157; 175; 195; 323; 327.” Zarząd Dróg Miejskich,

Urząd m.st. Warszawy, 2020.

Mapbox. “Mapbox Studio.” Generated September 2023. Accessed Feb-

ruary 2, 2024, at https://www.mapbox.com/about/maps.

My Journey: Helping Hampshire Get Around. “Southampton School

Streets programme.” Southampton City Council. 2021. Accessed

April 10, 2022, at https://myjourneyhampshire.com/education/

school-streets/southampton-school-streets-programme/.

To some extent the failure in this case could be partial-

ly justied by the negative assessment of viability of the

Bonaly Primary School for the implementation of a school

street closure program (Lawrence, Murrell 2016). Due to

the limited training sample, the algorithm was trained to

always generate a closure zone, even if the contexts are not

favourable. This hard requirement sometimes leads to in-

correct proposals. As the tool was being developed, a large

number of school street programs were implemented as

a result of the COVID-19 pandemic (Clarke 2022). Cur-

rently, a much larger dataset of successful closures can be

compiled to train the next iteration of the algorithm. Anoth-

er large problem arises from the incompleteness of the ana-

lysed contexts and the low resolution of the accessible data.

The utilized feature maps do not fully reect the real-world

complexities of implementing school street closures. While

the tool could denitely benet from more extensive, high-

resolution data, such data is not readily available or ma-

chine-friendly. These extended contexts could include:

– trac analysis, such as the road safety audit, recorded

incidents aecting the school community and trac inten-

sity measurements,

–

communication habits of the students and their parents,

– functional audit,

– database of stakeholders aected by the closure and

their characteristics,

– air quality and pollution measurements,

– records of other programs related to school street clo-

sures at the candidate school, including physical activity

encouragement projects, play streets, school gardening ini-

tiatives, local community activization, etc.

A more comprehensive data collection and integration

could improve the tool’s eectiveness and would proba-

bly increase the capabilities of the system. However, an

increase in the number of compiled context sources would

also reduce the applicability of the algorithm only to the

areas, which have these contexts recorded and accessible.

The current version of the algorithm can be applied to any

school that is represented on OpenStreetMap. Future re-

search should focus on expanding the dataset and improv-

ing the algorithm’s adaptability to diverse urban contexts.

Additionally, a collaborative approach involving stake-

holders from various sectors, including education, trans-

portation, and public health, is crucial for the successful

implementation and scaling of school street programs. By

doing so, cities can create safer, healthier and more vibrant

urban spaces that prioritize the well-being of children and

the entire urban community.

Conclusions

In conclusion, this study demonstrates the potential of

data-driven approaches to enhance the planning and im-

plementation of school street programs. The integration of

machine learning tools can streamline the selection pro-

cess and improve the design of these interventions, making

them more eective and context-sensitive. Policymakers

and urban planners should consider investing in the de-

velopment and deployment of such tools to support their

urban mobility goals.

146 Tomasz Dzieduszyński, Olga Czeranowska-Panufnik

Noble, Ruth, Naomi Boal, Chelle Watson, Emily Jackson, and Jack Po-

well. School Streets Intervention Sites vs. Control Sites Full Report.

Transport for London, 2021. Accessed June 20, 2022, at https://con-

tent.t.gov.uk/school-streets-evaluation-report-website.pdf

OpenStreetMap Contributors. “OpenStreetMap.” 2024. Accessed Janu-

ary 5, 2024, at https://www.openstreetmap.org/.

Perry, Clarence Arthur. “The Neighborhood Unit, a Scheme of Ar-

rangement for the Family-Life Community.” In Neighborhood and

Community Planning, Regional Plan of New York and Its Environs,

edited by Thomas Adams, Robert W. de Forest, John H. Finley

et al. Committee on Regional Plan of New York and Its Environs,

1929.

Perth and Kinross Council. “School Exclusion Zones.” Accessed June 20,

2022, at https://www.pkc.gov.uk/article/17611/School-exclusion-zones.

Pressl, Robert. “Bolzano: School Streets.” Eltis Mobility Portal. 2011.

Video, 5 min, 2 sec. Accessed April 2022. https://www.youtube.com/ -

watch?v=j8xKUjzaK8c.

Reeves, Will R. “Report of Committee on Street Play.” The Journal of

Educational Sociology 4, no. 10 (1931): 607–18.

Ritchie, Cameron. “Four More East Lothian Schools Set to Ban Trac from

Nearby Streets at Pick-up and Drop-o Times.” East Lothian Courier,

January 18, 2016. https://www.eastlothiancourier.com/news/14211155.

four-more-east-lothian-schools-set-to-ban-trac-from-nearby-streets-

at-pick-up-and-drop-o-times/.

Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. “U-Net: Convolu-

tional Networks for Biomedical Image Segmentation.” In Medical Im-

age Computing and ComputerAssisted Intervention – MICCAI 2015.

Lecture Notes in Computer Science, edited by Nassir Navab, Joachim

Hornegger, William M. Wells and Alejandro F. Frangi. Vol. 9351.

Springer, 2015. https://doi.org/10.1007/978-3-319-24574-4_28.

School Street Initiative. “What Are School Streets.” Published January,

2019. Accessed April 10, 2022, at http://schoolstreets.org.uk/.

Solihull Metropolitan Borough Council. “Solihull School Streets.” Ac-

cessed April 10, 2021, at https://www.solihull.gov.uk/Roads-pave-

ments-and-streetcare/Solihull-school-streets.

Stubbs, Alan. “School Streets – Trac Prohibition TROs” – An Assess-

ment Policy.” East Lothian Council, 2016. Accessed April 10, 2022,

at https://www.eastlothian.gov.uk/download/meetings/id/17220/04_

school_streets_trac_prohibition_tros_an_assessment_policy.

Urząd m.st. Warszawy. Bezpieczniej w drodze do szkoły. Startują kolejne

“Szkolne ulice.” Published September 20, 2022. Accessed Novem-

ber 12, 2022, at https://um.warszawa.pl/waw/strategia/-/startuja-

kolejne-szkolne-ulice.

Zarząd Dróg Miejskich [Warszawa]. “Droga na Szóstkę. Program po-

prawy bezpieczeństwa w okolicach szkół. Edycja I. Podsumowa-

nie audytów.” Published June 9, 2021. Accessed April 10, 2022, at

https://zdm.waw.pl/wp-content/uploads/2021/06/Droga-na-6_pod-

sumowanie_szczegolowe.pdf.

Streszczenie

Wprowadzanie ulic szkolnych: perspektywa uczenia maszynowego

Tematem artykułu są ulice szkolne, a zwłaszcza proces ich tworzenia z perspektywy uczenia maszynowego. Autorzy przedstawili w nim możli-

wości zastosowania generatywnych sieci przeciwstawnych typu GAN we wdrażaniu koncepcji zamknięcia ulic szkolnych, której celem jest poprawa

bezpieczeństwa i zmniejszenie natężenia ruchu wokół szkół. Przeanalizowali programy ulic szkolnych na całym świecie, identykując powtarzające

się wyzwania i proponując rozwiązania. Na podstawie analizy 51 skutecznych wdrożeń oraz nowej metody ekstrakcji cech miejskich opracowali

model uczenia maszynowego, który wspomaga wybór potencjalnych lokalizacji ulic szkolnych. Tak przygotowane narzędzie ma służyć usprawnie-

niu procesu wyboru i zwiększeniu efektywności projektu przez lepsze dopasowanie do lokalnego kontekstu. Pomimo pewnych ograniczeń, takich

jak trudność w odwzorowaniu wszystkich kontekstów przestrzennych, system dostarcza wartościowych wniosków dotyczących regulacji ruchu

miejskiego. Badanie wypełnia lukę w literaturze naukowej i oferuje podejście oparte na danych do projektowania bezpieczniejszych przestrzeni

miejskich.

Słowa kluczowe: ulice szkolne, uczenie maszynowe, smart city, generatywne sieci przeciwstawne, projektowanie architektoniczne wspomagane

komputerowo