Smart Mobility Support for Vehicle-based Tourism: Theoretical and

Technological Foundations

Sergey Mikhailov

1 a

, Alexey Kashevnik

1 b

, Alexander Smirnov

1 c

and Vladimir Parfenov

2 d

St. Petersburg Federal Research Center of the Russian Academy of Sciences (SPC RAS),

39, 14 Line, St. Petersburg, Russia Federation

ITMO University, Russia Federation

Keywords:

Smart Mobility, Tourism, Attraction Routes.

Abstract:

Vehicle-based tourism becomes more and more important in the era of the pandemic. Tourism management

is an important challenge for the tourist region development. The construction of a personalized attraction

visiting route for tourists with personal vehicles has a great impact on the tourist ﬂows. The authors propose

theoretical and technological foundations for smart mobility support of vehicle-based tourists. We propose

to predict tourist preferences by using deep neural networks for the prediction model’s implementation and

demonstrated 70-80% accuracy in training on completed tourist trips to St. Petersburg, Russia. The tourist

route attractiveness prediction was used to assess the constructed route quality. The attraction attractiveness

and attendance prediction together with potential tourist trajectory prediction were used for attraction selection

process personiﬁcation. The obtained results can be used in smart mobility support systems to improve the

travel experience.

1 INTRODUCTION

The tourism industry has grown rapidly in recent

years and has been intensively integrated with mod-

ern information and communication technologies.

Around 1.5 billion international tourist arrivals were

recorded worldwide in 2019, according to reports

from the World Tourism Organization (UNTWO).

Before the coronavirus disease (COVID-19) outbreak

restrictions applied, international travel was expected

to increase by 3.3% per year between 2010 and 2030.

It should be noted that the methods used by coun-

tries to combat COVID-19 such as population vac-

cination make perspectives on tourism growth in the

nearest future. Smart tourism services usage for road

travel should help to rebuild the tourist area around

the world in the shortest possible time (Bulchand-

Gidumal, 2022).

The fusion of information technologies and

tourism has given rise to the phenomenon of smart

tourism. Collection and analysis of data extracted

https://orcid.org/0000-0002-3738-0639

https://orcid.org/0000-0001-6503-1447

https://orcid.org/0000-0001-8364-073X

https://orcid.org/0000-0002-6139-9786

from various sources are typical activities for smart

tourism. They in combination with the use of ad-

vanced information technology make the tourist expe-

rience more enriching, efﬁcient, and sustainable. Re-

searches show that the use of various electronic ser-

vices and tools such as recommendation services, so-

lutions for building a route for visiting attractions can

improve the overall tourist experience from travel.

More and more tourists are using smartphones

during their trips that affects to attraction route for-

mation (Kim et al., 2021). In addition, the role of

vehicle-based tourism is also increasing (Cohen and

Hopkins, 2019). The amount of user-generated con-

tent generated during journeys is increasing from year

to year. Scientists use Big Data methods and tech-

niques to process large amounts of information, that

can be set as the basis for a predicting tourist be-

haviour model (Zhu and Shang, 2021). These mod-

els can be used to improve the overall performance of

various e-tourism services.

The paper presented theoretical and technologi-

cal foundations for smart mobility support of vehicle-

based tourists that includes the route conﬁguration

process for vehicle-based tourism with personalized

attraction selection and fast path construction be-

tween selected points of interest (POI). The main

Mikhailov, S., Kashevnik, A., Smirnov, A. and Parfenov, V.

Smart Mobility Support for Vehicle-based Tourism: Theoretical and Technological Foundations.

DOI: 10.5220/0011074100003191

In Proceedings of the 8th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2022), pages 105-115

ISBN: 978-989-758-573-9; ISSN: 2184-495X

105

idea is to form attraction set and then propose to the

tourist an effective attraction visiting plan. We use the

tourist preferences prediction models for improving

the tourist route quality by considering both context

and historical information. The prediction models are

based on the neural network and open-source data

usage for tourist behaviour patterns extraction. We

propose to use Kohonen self-organizing maps, EM-

algorithm (Expectation Maximization), and deep neu-

ral networks.

The paper is structured as follows. The Section 2

presents related work on the topic of tourist routes

construction. Section 3 describes the proposed ap-

proach to individual route conﬁgurations based on

context, historical data, and tourist preferences. Sec-

tion 4 describes the presented approach implementa-

tion and evaluation. Section 5 provides a paper sum-

mary.

2 RELATED WORK

(Chen et al., 2020) presents the DCC-PersIRE

method, which determines the interests of the user

and recommends an individual route based on them.

To extract meta-information from text descriptions

of places of interest the authors suggest using deep

machine learning models with unsupervised learning.

The route is built based on an iterative local search.

(Malik and Kim, 2019) propose a method for gen-

erating the optimal tourist route. Algorithms used in

the proposed methodology are neural networks for

prediction and particle as well as swarm optimiza-

tion to ﬁnd the optimal route. The authors develop

an objective function for route optimization based on

ﬁve route parameters: distance, trafﬁc congestion,

weather conditions, route popularity, and user pref-

erences.

(Tsai et al., 2019) offers a way to recommend

POI based on the photo analysis from social networks.

The obtained GEO information is clustered into vari-

ous categories using the topic modelling approach of

latent Dirichlet distribution. Based on the obtained

clusters tours are generated using the LSTM network.

(Zheng and Liao, 2019) consider the problem of

tourist route conﬁguration among groups of hetero-

geneous tourists using the Pareto optimality criterion.

To solve this problem they proposed to use the ant

colony algorithm for routing among attractions and a

differential evolution algorithm for generating sets of

attractions.

(Mukhina et al., 2018) uses various social net-

works to form an assessment of the attractiveness of a

place and also uses knowledge of the region’s popular

types of attractions as context when building a route

and performs simulation events when managing the

route conﬁguration.

(Taylor et al., 2018) uses linear programming

algorithms to calculate a set of attractions recom-

mended for a tourist to visit. The algorithm presented

by the authors selects attractions located near the ho-

tels where tourists stay.

(Hti and Desarkar, 2018) use the location taken

from social networks to generate recommendations

for visiting attractions. Social networks act as a

source of information about visited places and reac-

tions to them. Attractions are subjected to two-level

ﬁltering. The distance between the remaining routes

is built using the Floyd-Warshall algorithm.

(Bartie et al., 2018) describes the SpaceBook

project that implements an idea of a virtual guide

driven by tourists based on voice analysis. The guide

can notify a tourist of nearby points of interest on an

interactive map. Attraction information is collected

from open sources such as OpenStreetMap and social

networks.

(Santos et al., 2017) presented a hybrid recom-

mendation system that builds a tourist route based on

user proﬁles with disabilities. The paper proposes to

use an ontological approach to modelling knowledge

in the tourism topic. Recommendations are based

on attractions categories, the potential emotional in-

volvement of tourists, and accessibility and amenities

for travellers with disabilities.

(Chen and Tsai, 2017) describes the development

of a personalized and location-based mobile travel ap-

plication. The application is based on the iBike sys-

tem in Taichung City, Taiwan and uses a hybrid ﬁlter-

ing technique to collect travel information. The de-

veloped system adapts the ant colony algorithm for

its work to ﬁne-tune geolocation recommendations to

tourists. Authors used a technology adoption model

to interpret the adoption of information technology

by users. It includes three evaluation criteria: per-

ceived ease of use, perceived usefulness, and use of

behavioural intent. The information systems success

evaluation model contains six factors: system qual-

ity, information quality, system usage, user satisfac-

tion, individual user impact, and organizational im-

pact. These factors are related to the evaluation of the

success of the information system.

(Colomo-Palacios et al., 2017) described the

POST-VIA 360 platform designed to analyze the full

life cycle of tourist loyalty after the ﬁrst visit to the

region. Based on the carried out analysis the platform

can offer recommendations for visiting new places

based on tourist location and artiﬁcial immunity prin-

ciples.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

106

(Kotiloglu et al., 2017) suggests a “Filtering then

generating a tour” approach for creating personalized

recommendations on tourist routes based on infor-

mation from social networks and other online data

sources. The authors proposed to apply collaborative

ﬁltering to deﬁne a subset of additional points of inter-

est that maximize the user’s potential satisfaction with

the route while the itinerary must select mandatory

places that tourists must visit. The main orientation

problem is solved using the iterative search for prohi-

bitions algorithm, which must create travel itineraries.

These itineraries contain all the required places and

maximize the overall tourist satisfaction when using

additional daily visited places, taking into account ac-

cess times, opening hours, restrictions on the tourist

time, his/her budget, etc.

(Cenamor et al., 2017) presented the PlanTour

system, which creates personalized tourist routes us-

ing human-generated information collected from the

travel social network MINUBE. The system follows

an automated planning approach to create a multi-day

plan with the most interesting sights of the visited re-

gion. In particular, the system collects information

about users and points of interest from MINUBE and

groups these points using clustering methods to di-

vide the problem into subtasks. Then the tourist uses a

destination-independent automated planner that ﬁnds

quality travel plans. According to the authors of the

paper, unlike other tourist recommendation systems,

the PlanTour planner can organize relevant points of

interest based on the user’s expected trips and user

ratings from a real social network.

(Nilashi et al., 2017) authors propose a proto-

type of a travel advisory system that offers a method

based on multicriteria collaborative ﬁltering. This

method improves the prediction accuracy of tourism

recommendation systems by using clustering, sam-

ple size reduction, and forecasting methods. The au-

thors used an adaptive network based on a fuzzy in-

ference system and support vector machines for pre-

dictions. Principal component analysis was used to

reduce the dimensionality of the samples, Kohonen

self-organizing maps and EM-algorithm (Expectation

Maximization) were chosen as well-known clustering

methods. To improve the accuracy of recommenda-

tions of multicriteria collaborative ﬁltering, the clus-

ter ensemble approach, Hypergraph Partitioning Al-

gorithm (HGPA) have been applied to SOM and EM

clustering results. The authors evaluated the accuracy

of the recommendation method on the TripAdvisor

dataset and their experiments indicate that cluster en-

sembles are more predictive than single-criteria clus-

tering techniques.

(Gavalas et al., 2017) presented a mobile travel

guide, which can correctly process points of interest

in their original form when generating routes without

converting the geometry to a point. The routes created

by the guide include extensive walking areas. Route

building is based on local search with iterations. User

preferences and the total time to complete the entire

tour are taken as contextual information.

Based on the analysis of modern routes conﬁgura-

tion models we proposed to divide the conﬁguration

management task into two main subtasks: formation

of a set of attractions and the creation of a visiting se-

quence for a set of attractions. Existing approaches

actively use technologies for building recommenda-

tions, to form sets of attractions, and actively person-

alize the route based on the restrictions imposed by

the tourist. However, the considered works do not

fully work with the context, and also do not use the

historical data of tourists and the region when build-

ing routes.

3 THEORETICAL AND

TECHNOLOGICAL

FOUNDATIONS

We proposed theoretical and technological founda-

tions for smart mobility support of vehicle-based

tourists. For the tourist support we propose to form

attraction set and based on this set make the visit-

ing plan for the tourist based on his/her preferences as

well as context situation in the region (see Figure 1)

In the Section we consider the route conﬁguration,

tourist preferences prediction, attraction set formation

as well as attraction visiting plan creation. The con-

sidered tasks allows to support of the smart mobility

support of the vehicle based tourist by proposing to

him/her efﬁcient attraction attending plan taking into

account his/her preferences and current situation in

the region.

3.1 Route Conﬁguration

The route conﬁguration process within tourist region

T consists of two part: attraction set formation

seq

(Tr,C

) = {(A

),(A

),...,(A

)}

and attractions visiting sequence creation

R(A

seq

,Tr,C

), where Tr — tourist in-

formation, A

— region attraction, S

— personalized

attraction score for tourist, C

— tourist region

context, C

— tourist context and C

— region

attractions context. The route quality is measured

by Equation 1, where f

dist

— route distance rating,

Smart Mobility Support for Vehicle-based Tourism: Theoretical and Technological Foundations

107

time

— route time rating, f

pred

— route attrac-

tiveness rating, f

bud

— tourist spending rating and

α,β,γ,ζ — correction coefﬁcients.

score

= α f

dist

(R,C

) + β f

time

(R,C

)

+γ f

pred

(R,Tr,C

) + ζ f

bud

(R,C

)

(1)

Route distance and time rating functions com-

pares (Equation 2) route characteristics (D

, T

—

route distance and time) with “ideal” route (R

min

— minimal distance and time), which

which is built without taking into region con-

text. The Route attractiveness rating function (Equa-

tion 3, Sc — tourist subjective route score, [Sc

min

1,Sc

max

= 5] — route estimate limits) offers an

assessment of how much the tourist will like the

proposed route and tourist spending rating function

(Equation 3, budget

— planned tourist budget,

cur

— route costs, bud

— attraction entry fee)

checks if the tourist has exceeded the planned spend-

ing.

dist

(R,C

) =







min

, D

> D

min

1, D

≤ D

min

∈ R, D

min

∈ C

time

(R,C

) =







min

, T

> T

min

1, T

≤ T

min

∈ R, T

min

∈ C

(2)

pred

(R,Tr,C

) =











1, Sc = Sc

max

0, Sc = Sc

min

Sc − 1

max

− 1

, otherwise

Sc ∈ R

bud

((R,C

) =











0 budget

< B

cur

0.5 budget

= B

cur

1, budget

> B

cur

∑

i=m

bud

(3)

The following restrictions is taken into con-

sideration in route construction process: T

≤

des

,Budget

≥ B

cur

,...,A

∈ A

seq

,...,A

/∈

seq

,α + β + γ + ζ = 1. The main goal of the pro-

posed approach is to increase tourist satisfaction by

maximizing f

score

3.2 Tourist Preference Prediction

Figure 2 presents route conﬁguration management

scheme and highlights tourist activity managing tasks.

We propose to track the movement of tourists based

on GPS, extract data on reviews, and ratings of at-

tractions, as well as indirectly receive information

about the attractions visiting by using a smartphone.

Based on data sources the following tasks were iden-

tiﬁed: identifying tourists behaviour groups based on

preferences similarity, assessing tourists satisfaction

after passing the formed routes, assessing changes

in tourist ﬂows and identifying typical routes among

tourists.

3.3 Attraction Set Formation

Attraction set formation scheme is presented in Fig-

ure 3. Recommendation system based on synthetic

coordinates Srcs (Papadakis et al., 2017) takes the

tourists and experts attraction scores as input and pre-

dicts the personalized non-visited attraction ratings

scored

for the speciﬁc tourist. The recommendation

creation is based on the Vivaldi algorithm (Moravek

et al., 2011), which simulates a network of physical

springs, placing imaginary springs between pairs of

network nodes such as tourists and attractions. The

used algorithm is not parameterized, does not require

ﬁne-tuning and is more resistant to the “cold start”

problem.

At the same time, an attraction attractiveness pre-

diction model ANN

clust

is applied to change tourist at-

traction ratings. These models take as input the aggre-

gated information about tourist actions and performed

routes for a certain period and determine the tourist

cluster, which represents as “behavioural” group. For

each cluster, the most popular and high rated visited

attractions were gathered and their ratings were in-

creased by 0.2 to increase the probability of further

selection by Srcs.

After A

scored

acquiring the personalized attrac-

tion list is ﬁltered by the following metrics: re-

moval of undesirable for visiting places, allocated by

the tourist himself Rt and removal previously visited

places which have been recommended by approach

vis

. Then two tourist preferences prediction mod-

els are used to reﬂect the region historical data. The

potential route prediction model ANN

pred tra j

recon-

structs the most popular route among the tourists by

analyzing previous trips trajectories. Based on the ob-

tained route the model retrieves the most popular POI

lying on the constructed path and increase their rating

by 0.1. The attraction attendance prediction model

ANN

pred att

compares the predicted attraction atten-

dance with the average attendance values provided by

the region. If the difference is positive, the model

reduces the attraction rating by 0.3, otherwise it in-

creases it.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

108

1. Attraction set formation

2. Attractions visiting plan

Personal attractions ratings selection

Attraction sets formation

Personal attractions ratings

Route generationRoute evaluation

• Attraction attractiveness prediction

• Attraction attendance prediction

• Potential routes prediction

Attraction sets

Generated routes

• Route attractiveness prediction

Tourist

restrictions

Best POI

route

Figure 1: Smart mobility support of vehicle-based tourists.

Route con gura on management

A rac ons set forma on

A rac on route forma on

Tourist

preferences

Expert ra ngs

Context &

historical data

Context &

historical data

Region

restric ons

Tourist

restric ons

Assessment of tourist sa sfac on

a er formed routes passing

Tourist route templates

iden ca on

Similar preferences

iden ca on between

tourists

Tourist ow changes analysis

aﬀected by formed routes

A rac on ra ngs & reviews

Trip ra ngs

Smartphone

sensors (GPS)

A rac on

endance

Figure 2: Route conﬁguration management scheme.

The ﬁltered attraction list A

f iltered

is divided into

the n attraction sets A

seq

,...,A

seq

by using the slid-

ing windows technique with length k, which is set by

tourist restrictions Rt. The tourist starting and ﬁnish-

ing points are inserted into each attraction set. All

constructed sets are validated on the simpliﬁed region

graph, where the vertex is a regional attraction and

the edge is a minimal distance between two POI. The

validation consists of potential route duration compar-

ison with the desired duration given by tourist and at-

traction working hours check. If the number of val-

idated sets is less than m, the cycle of attraction sets

creation and validation continues. In the end the ap-

proach returns the personalized and validated attrac-

tion sets A

seq

,...,A

seq

3.4 Attractions Visiting Plan

Figure 4 describes attractions visiting plan creation

process. At ﬁrst the full region graph G is constructed,

based on the information from Openstreetmap

. In

addition, the road trafﬁc information is gathered from

the smart city services and applied to G for creating

https://www.openstreetmap.org

Smart Mobility Support for Vehicle-based Tourism: Theoretical and Technological Foundations

109

n sets with k attractions

Correct sets

New attraction set generation

Territorial

restrictions

application

Visited

attractions

fjltration

Must-see attraction

inclusion

Data ANN (Predictive models) Computing models

Result data passing

Additional information usage

Figure 3: Attraction Set Formation.

a graph G

mod

with current region state information.

Next, the following steps are cyclically performed.

• The m attraction sets are taken as input for se-

quence creations. attractions inside the sets are

sorted by proximity to the tourist.

• For each set the route R

is created by using

the multilevel Djikstra algorithm (Delling et al.,

2017). This modiﬁcation of Djikstra algorithm

allows to quickly build routes for the proposed

places due to the internal representation of the

graph in the form of nested cells.

• For each constructed route R

the quality rating

score

is computed by using the f

score

. The f

score

uses 0.25 as a default value for α, β, γ,ζ correction

coefﬁcients if the tourist did not enter his prefer-

ences for the formation of routes in the mobile ap-

plication, otherwise, the different distribution of

weights is applied. The route attractiveness pre-

diction model is used as part of f

score

• The quality scores and route characteristics are

saved in the tabu list, and the mutation process

is occurred by exchanging 20% of attractions be-

tween random set pairs. If the received route was

previously evaluated in the tabu list, the route is

generated again.

The condition for terminating the cycle is exceed-

ing the iteration limit (it is conﬁgured individually for

each region) or exhausting the set of routes available

for generation. In the end, the tourist receives the

route with the maximum R

score

, which represents the

highest quality available route.

3.5 Tourist Preferences Prediction

Models

Based on the examination of the modern studies with

the analysis of tourists states and a tourist region

within the framework of the predictive models use,

we decided to use solutions based on a neural network

approach. Neural network models have a better pre-

dictive ability for identifying non-obvious functional

dependencies within the tourism system than tradi-

tional statistical analysis methods. Another advantage

of such models is independence on the speciﬁc statis-

tical characteristics of the available dataset, as well as

resistance to incomplete, redundant and noisy data. It

is also worth noting that neural network models show

the best results for predicting the states of objects and

subjects of systems when accessing a large amount of

historical data that a tourist system can provide. Neu-

ral network models meet the requirements of taking

into account seasonality and delays in assessing man-

agement efﬁciency through the use of historical data.

In Table 1 we proposed the several types of neu-

ral networks to solve different prediction tasks. We

taking into account the characteristics of datasets for

each forecasting task. Predicting the attractiveness

of the route and attractions do not have temporary

events, therefore the type of networks, based on the

LSTM approach and taking into account time events

are not suitable for this type of task, in contrast to

the tasks of predicting the attendance of attractions

and potential routes. Predicting the attractiveness of a

route is solved within the framework of classiﬁcation

problems, for which a deep neural network can be ap-

plied. For the task of predicting the POI attractive-

ness, self-organizing Kohonen maps are used, which

solve the problem of clustering users.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

110

Sorted by tourist proximity

Rating saving, attraction set modification

Cycle X times

Figure 4: Attractions visiting sequence creation.

Table 1: Neural network architecture selection based on prediction task.

Task Chosen architec-

ture

Dataset information Network type justiﬁcation

Route attractive-

ness prediction

Deep neural net-

work

Routes ratings by tourists,

routes characteristics

Supervised learning, trips do

not affect each other, no time

events

Attraction attrac-

tiveness prediction

Self-organizing

maps

Average characteristics of com-

pleted routes, tourist prefer-

ences

Unsupervised Learning, no

time events

Attraction atten-

dance prediction

Bidirectional

LSTM

Time distribution of tourist at-

tractions visitation, weather and

budget costs

Supervised learning, events are

distributed over time and has in-

ﬂuence on each other, context

clariﬁes the situation with at-

tendance

Potential route pre-

diction

Bidirectional

LSTM

GPS trajectories of tourists Supervised learning, events are

distributed over time and inﬂu-

ence each other, time and type

of terrain set the context

The proposed neural network architecture consists

of one input layer, one normalisation layer, three hid-

den layers with 128 neurons, and one output layer.

Hidden layers used ReLU activation function.

Implementation details of tourist preferences pre-

diction models described in the (Mikhailov and Ka-

shevnik, 2021), (Mikhailov and Kashevnik, 2020).

The overall accuracy of model prediction was 70-80%

depending on the prediction type. For the route at-

tractiveness prediction model, as a context expected

distance and duration of the formed route the num-

ber of visited attractions and their average rating, cur-

rent trafﬁc assessment, current weather situation and

tourist attractions preferences are used. As for at-

traction attractiveness prediction, the model deﬁnes

the context as the attendance of the attractions of the

region, the number of generated tourist content dur-

ing tourist trips, the average budget spend and tourist

preferences. Historical data is determined by the aver-

age duration and distance of trips, the number and rat-

ing of points of interest, and the average travel speed.

For attraction attendance prediction information

about the activity of visiting attractions with a break-

down by day is used as historical data. The weather

situation on a speciﬁc day and the total spending of

the monetary budget is used as contextual informa-

tion. The model for predicting potential routes based

on the days of the week and the type of road can form

potential tourist routes that can be used as an addi-

tional source of extraction of popular attractions. The

neural network model uses a part of the path trajec-

tory in the format (azimuth between points, distance

between points) as historical data.

Smart Mobility Support for Vehicle-based Tourism: Theoretical and Technological Foundations

111

4 IMPLEMENTATION AND

EVALUATION

4.1 Architecture of the Protopyte

Figure 5 describes the system prototype architecture,

which implements the proposed route construction

approach. The prototype uses a data-driven approach

for route conﬁguration creation and consists of differ-

ent micro-services. For the functioning of services in

an isolated environment, the Docker

was used, which

allows store each of the system components in a con-

tainer. This virtualization technique allows to conve-

niently deploy, maintain and scale a system for differ-

ent tourist regions.

The system prototype architecture can be divided

into three parts: collecting data about a tourist when

moving in a tourist region, tourist behaviour analy-

sis and the tourist route creation, visualization of the

results of the analysis of tourist behaviour. During

the travel a tourist uses his electronic device (smart-

phone, tablet, etc.) with a tourist support mobile ap-

plication (Mikhailov and Kashevnik, 2018). This ap-

plication monitors the state of sensors (GPS, magne-

tometer, etc.), shows the route among the attractions,

offers to evaluate the route and attractions. The appli-

cation also sends information about the tourist actions

and statistics from sensors to the database.

The tourist information transfer is carried out

over the HTTP protocol to the REST-API service.

The entry point of this service is the Nginx HTTP

server

, which provides access to a backend written

in the Python programming language. Due to a large

amount of information coming from tourists in the

tourist region, it was decided to use asynchronous

framework aiohttp

for building a REST-API service

that allows processing a large number of requests in

parallel without high costs for I/O operations. Also,

to ensure parallelization of requests, and Gunicorn

HTTP server was used, which allows to launch and

manage parallel instances of the tourist information

processing service.

A Postgresql

database was used as the informa-

tion storage. All information is stored with indexed

timestamps according to the concept of a digital pat-

tern of life (Mikhailov and Kashevnik, 2020). This

enable fast information analysis about the tourist and

the tourist region. Elasticsearch technology stack was

used as additional storage of unstructured data —

https://www.docker.com/

https://nginx.org/

https://docs.aiohttp.org

https://www.postgresql.org/

Logstash — Kibana

, which allows to track incoming

tourist information in real-time.

Tourist behaviour analysis services use deep

ANNs, available information from Postgresql and

Elasticsearch databases to extract the behavioural

characteristics of tourists and use them to support

vehicle-based tourist activity. The region informa-

tion extraction service (Smirnov et al., 2020) retrieves

information about attractions within a tourist region

based on geo-information from the OpenStreetMap

service, information from Wikipedia (multimedia in-

formation) and Google Places (attraction rating in-

formation). Deep neural network models are imple-

mented using the Python programming language and

the Tensorﬂow

library. The formation of sets of

attractions uses in its work the recommending sys-

tem ScoR (Papadakis et al., 2017), based on the use

of synthetic coordinates. The OSMR

platform was

used to create a sequence of attraction visitation to

implement a multilevel Dijkstra algorithm.

The service for visualizing the results of the anal-

ysis of tourist behaviour was implemented using the

Javascript programming language and the Vue.js li-

brary

, which allows to implementation of a Single

Page Application (SPA) approach to creating web-

sites. The service allows to view current tourist trips

in real-time, view the state of the tourist and the

tourist region as well as visualize the results of pre-

dictive models.

4.2 Experiments

The approach evaluation consisted of a “blind” com-

parison of the expert routes and the generated routes

by the described approach. Each of the proposed

routes passed through the city of St. Petersburg, Rus-

sia. Experiment participants took part in a survey

based on a website that displayed routes on an interac-

tive map. The website was developed in the Javascript

programming language using the Vue.js library, and

the results were processed using the Python program-

ming language.

At the survey, start participant had to enter per-

sonal information — gender, age and information

about St. Petersburg residence. 63 people partici-

pated in the survey in total, 31 people were aged 18–

30 years (49.2% of the total sample), 18 people —

at the age of 31–45 years (28.6%), 14 people — at

the age of 46–64 years (22.2%). Most of the subjects

https://www.elastic.co/what-is/elk-stack

https://www.tensorﬂow.org/

http://project-osrm.org/

https://vuejs.org/

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

112

Gunicorn (Docker)

Docker

Gunicorn (Docker)

Docker

Tourist 1

Sensors

...

Tourist N

Sensors

Aioh p

instance

Aioh p

instance

...

Nginx

Postgresql

Route data

Processed data

Aioh p

instance

Aioh p

instance

...

Vue.js

Nginx

Requested data

Analyst

Visualiza on request

Visualiza on

Logs

Smartphone

(client app)

Smartphone

(client app)

OSMR

A rac on set order forma on

(Python) - service

SCOR

A rac on set construc on

(Python) - service

Analyze of tourist behavior

(Python) – server app

Region informa on extrac on (Python) - service

Predic ve ANN

(Python + TensorFlow)

Google Places

Analyzes results + routes

OpenStreetMap

Wikipedia

Region and a rac ons data

A rac ons set

Order of a rac ons to visit

Generated route

Figure 5: System prototype architecture model.

live in St. Petersburg (39 people — 61.9% of the to-

tal sample), the remaining 24 (38.1%) people lived in

the Novgorod the Great, Kaliningrad and Moscow re-

gions.

The survey participants were asked to evaluate

ﬁve routes groups (Fig. 6), consisting of 3–7 entries.

Route groups could be either general or thematic. The

general group contains less than 50% popular region

points of interest within a given attraction set. The

thematic group contains more than 50% popular at-

tractions or all route POI were subject to the same

theme such as “Literature in St. Petersburg”, etc.

When building routes, the approach took into account

the current context and historical information about

the attractions and the region. The participants were

asked to rate the routes within each group on a ten-

point scale, where 10 is the highest possible route

quality score.

Each group contains one approach generated

route, at least one expert’s route and at least one

“mixed” route. The routes of GPSMyCity, Inspirock,

TripAdvisor services were used as expert routes

sources. The “mixed” routes has attractions set based

on the expert’s route, however, the routing process

between POI was done using the Multilevel Djikstra

algorithm. Each route was obtained in audio guide

form, which implies not direct visiting attractions and

receiving multimedia information about them using a

tourist smartphone.

As part of the ﬁrst group of routes, the goal was

to compare the routes of attractions in the centre of

St. Petersburg, located near the Hermitage. As part of

the second group of routes, sights associated with the

Hermitage, the Peter and Paul Fortress and the Cruiser

Aurora were used. The third group of routes consid-

ered the sights of the Vasileostrovsky district, but at

the same time, when constructing routes, the situation

was stimulated, when several popular attractions were

packed. In the fourth and ﬁfth groups, routes built in

the area of Apraksin Dvor and Admiralteysky Island

were compared.

For each of the routes within the groups, the arith-

metic mean of all ratings was calculated (Table 2). As

a result we made the following conclusions:

• Participants preferred the generated routes over

experts ones in the general route groups. How-

ever, in the thematic group’s participants rated the

expert’s routes higher. This is due to the general

decrease in the number of popular attractions in

the approach routes and the consideration of POIs

attendance.

• An analysis of the participant comments showed

that other cities residents (10 out of 24) more often

than indigenous people (8 out of 39) noted an in-

teresting choice of attractions in the routes created

using the presented approach. Some survey par-

ticipants used the routes proposed by the method

and highly appreciated it in practice.

Smart Mobility Support for Vehicle-based Tourism: Theoretical and Technological Foundations

113

Figure 6: Survey page with interactive routes.

Table 2: Average route scores within groups.

# Approach routes Experts routes

1 7.83 7.58

2 8.23 6.32

3 7.12 7.83

4 6.88 5.67

5 7.43 6.58

5 CONCLUSIONS

The paper presented theoretical and technological

foundations for a tourist smart mobility support tak-

ing into account historical data and tourist prefer-

ences. The evaluation shows that the proposed ap-

proach is effective in the tourism domain and allow to

recommend relevant attractions to him/her. The im-

plemented survey shows that the overall route scores

increase to 8.5% when using the presented approach

that is in contrast to the routes proposed by experts,

takes into account both the context of the tourist and

the tourist region and work with the accumulated his-

torical data, which allows the personalizing route to

the tourist.

The presented approach has limitations. We do

not take into account time required to visit the attrac-

tion as well as opening hours. This ideas is our future

work.

ACKNOWLEDGEMENTS

Research has been supported by the Russian Foun-

dation for Basic Research. Attraction set formation

(Section 3.3) is due to the project #20-07-00490, at-

traction visiting plan construction (Section 3.4) is due

to the project #20-07-00455. The experiments pre-

sented in the Section 4 are due to Russian State Re-

search FFZF-2022-0005.

REFERENCES

Bartie, P., Mackaness, W., Lemon, O., Dalmas, T., Ja-

narthanam, S., Hill, R. L., Dickinson, A., and Liu, X.

VEHITS 2022 - 8th International Conference on Vehicle Technology and Intelligent Transport Systems

114

(2018). A dialogue based mobile virtual assistant for

tourists: The spacebook project. Computers, Environ-

ment and Urban Systems, 67:110–123.

Bulchand-Gidumal, J. (2022). Post-covid-19 recovery of is-

land tourism using a smart tourism destination frame-

work. Journal of Destination Marketing & Manage-

ment, 23:100689.

Cenamor, I., de la Rosa, T., N

nez, S., and Borrajo, D.

(2017). Planning for tourism routes using social net-

works. Expert Systems with Applications, 69:1–9.

Chen, C.-C. and Tsai, J.-L. (2017). Determinants of behav-

ioral intention to use the personalized location-based

mobile tourism application: An empirical study by in-

tegrating tam with issm. Future Generation Computer

Systems.

Chen, L., Zhang, L., Cao, S., Wu, Z., and Cao, J. (2020).

Personalized itinerary recommendation: Deep and

collaborative learning with textual information. Ex-

pert Systems with Applications, 144:113070.

Cohen, S. A. and Hopkins, D. (2019). Autonomous vehicles

and the future of urban tourism. Annals of Tourism

Research, 74:33–42.

Colomo-Palacios, R., Garc

ıa-Pe

nalvo, F. J., Stantchev, V.,

and Misra, S. (2017). Towards a social and context-

aware mobile recommendation system for tourism.

Pervasive and Mobile Computing, 38:505–515. Spe-

cial Issue IEEE International Conference on Pervasive

Computing and Communications (PerCom) 2016.

Delling, D., Goldberg, A., Pajor, T., and Werneck, R. F.

(2017). Customizable route planning in road net-

works. Transp. Sci., 51:566–591.

Gavalas, D., Kasapakis, V., Konstantopoulos, C., Pantziou,

G., and Vathis, N. (2017). Scenic route planning for

tourists. Personal and Ubiquitous Computing, 21.

Hti, R. and Desarkar, M. (2018). Personalized tourist pack-

age recommendation using graph based approach.

pages 257–262.

Kim, H., Koo, C., and Chung, N. (2021). The role of mobil-

ity apps in memorable tourism experiences of korean

tourists: Stress-coping theory perspective. Journal of

Hospitality and Tourism Management, 49:548–557.

Kotiloglu, S., Lappas, T., Pelechrinis, K., and Repoussis, P.

(2017). Personalized multi-period tour recommenda-

tions. Tourism Management, 62:76–88.

Malik and Kim (2019). Optimal travel route recommenda-

tion mechanism based on neural networks and particle

swarm optimization for efﬁcient tourism using tourist

vehicular data. Sustainability, 11:3357.

Mikhailov, S. and Kashevnik, A. (2018). Smartphone-based

tourist trip planning system: a context-based approach

to ofﬂine attraction recommendation. MATEC Web of

Conferences, 161:03026.

Mikhailov, S. and Kashevnik, A. (2020). Tourist behaviour

analysis based on digital pattern of life—an approach

and case study. Future Internet, 12:165.

Mikhailov, S. and Kashevnik, A. (2021). Car tourist tra-

jectory prediction based on bidirectional lstm neural

network. Electronics, 10(12).

Moravek, P., Komosny, D., Vajsar, P., Sveda, J., and Handl,

T. (2011). Study of vivaldi algorithm in energy con-

straint networks. Advances in Electrical and Elec-

tronic Engineering, 9.

Mukhina, K. D., Visheratin, A. A., and Nasonov, D. (2018).

Building city-scale walking itineraries using large

geospatial datasets. In 2018 23rd Conference of Open

Innovations Association (FRUCT), pages 261–267.

Nilashi, M., Bagherifard, K., Rahmani, M., and Rafe, V.

(2017). A recommender system for tourism industry

using cluster ensemble and prediction machine learn-

ing techniques. Computers & Industrial Engineering,

109:357 – 368.

Papadakis, H., Panagiotakis, C., and Fragopoulou, P.

(2017). Scor: A synthetic coordinate based recom-

mender system. Expert Systems with Applications, 79.

Santos, F., Almeida, A., Martins, C., Gonc¸alves, R., and

Martins, J. (2017). Using poi functionality and acces-

sibility levels for delivering personalized tourism rec-

ommendations. Computers, Environment and Urban

Systems, 77.

Smirnov, A., Kashevnik, A., Mikhailov, S., Shilov, N.,

Orlova, D., Gusikhin, O., and Martinez, H. (2020).

Context-driven tourist trip planning support sys-

tem: An approach and openstreetmap-based attrac-

tion database formation. In Popovich, V., Thill, J.-C.,

Schrenk, M., and Claramunt, C., editors, Information

Fusion and Intelligent Geographic Information Sys-

tems, pages 139–154, Cham. Springer International

Publishing.

Taylor, K., Lim, K. H., and Chan, J. (2018). Travel itinerary

recommendations with must-see points-of-interest. In

Companion Proceedings of the The Web Conference

2018, WWW ’18, page 1198–1205, Republic and

Canton of Geneva, CHE. International World Wide

Web Conferences Steering Committee.

Tsai, C.-Y., Paniagua, G., Chen, Y.-J., Lo, C.-C., and Yao,

L. (2019). Personalized tour recommender through

geotagged photo mining and lstm neural networks.

MATEC Web of Conferences, 292:01003.

Zheng, W. and Liao, Z. (2019). Using a heuristic approach

to design personalized tour routes for heterogeneous

tourist groups. Tourism Management, 72:313–325.

Zhu, W. and Shang, F. (2021). Rural smart tourism under

the background of internet plus. Ecological Informat-

ics, 65:101424.

Smart Mobility Support for Vehicle-based Tourism: Theoretical and Technological Foundations

115