Applying Activity-Based Models to Integrate Labeled Preset Key Events

in Intra-Day Human Mobility Scenarios

Patrik Gonc¸alves

1 a

and Harald Baier

2 b

Zentrale Stelle f

ur Informationstechnik im Sicherheitsbereich, Zamdorfer Straße 88, M

unchen, Germany

Research Institute Cyber Defence (CODE), Universit

at der Bundeswehr M

unchen,

Carl-Wery-Straße 22, Munchen, Germany

Keywords:

Activity-Based Models, Non-Typical Behavior, Preset, Custom, A-Priori, Activities, Key Events, Outliers,

Human Mobility, Dataset, Intra-Day.

Abstract:

The generation of synthetic human mobility scenarios is often realized through data-driven or rule-based

approaches. They work in a ﬁre-and-forget principle and provide limited support to induce controlled activities

in simulated scenarios. However, including controlled preset activities in the generation phase enables the

creation of mobility scenarios that include a-priori known outliers or key events. Such mobility test datasets

might be used in outlier detection for machine learning algorithms or for inducing non-typical mobility, where

models do not exist or are too complex to construct. In this work we propose an activity-based scheduler to

include controlled preset key events in the scheduling process of daily human mobility scenarios. Further, with

our rule-based approach we can synthesize new activities of a target region even when initial data is unavailable

or missing. In addition we propose a hierarchical methodology to iteratively add activities according to their

number of constraints and provide a publicly available Python-based implementation. Our validation shows

that our approach is able to integrate non-typical behavior in typical mobility scenarios.

1 INTRODUCTION

Contemporary mobile devices such as smartphones,

smart wearables, GPS navigation systems or Internet

of Things devices are able to determine a device’s

position in real time. By storing the device’s cur-

rent position over long time periods, human individ-

uals wearing or using these devices can be tracked in

a very convenient way. The resulting human-based

geospatio-temporal movement is referred to as hu-

man mobility (Barbosa et al., 2018; Castiglione et al.,

2015).

Geospatial datasets of human mobility are bound

to speciﬁc regions and describe a speciﬁc pattern ac-

cording to the used modes of transportation. There-

fore, the transferability to other regions and modes is

limited (Luca et al., 2021), hence a researcher needs

to gather or synthesize new data for the target re-

gion ﬁtting to his use case. Current data-driven ap-

proaches, e.g., machine learning (ML)-based solu-

tions like (B

osche et al., 2012; Drchal et al., 2019),

https://orcid.org/0000-0001-5561-8818

https://orcid.org/0000-0002-9254-6398

still need a representative dataset in the training phase,

which on the other hand is often acquired from mobil-

ity surveys and custom-acquired tracking data of the

target region. On the other hand, these datasets of-

ten suffer from weaknesses, which render their use

in practice difﬁcult to impossible. For instance, if a

dataset is anonymized, proprietary (e.g., with respect

to a closed format or high costs), unlabeled, unreal-

istic, unsuitable for the individual use case (i.e., the

dataset represents an incorrect mobility pattern for

the respect application) or in the worst case not com-

plete (Aschenbruck et al., 2010).

Applying human mobility models is a convenient

way to generate synthetic data on a large scale, where

the other options fail due to spatial restrictions (e.g.,

creating tracking data of a large area or collecting

data from a restricted area), resource limitations (e.g.,

limited in time or funding) or legal constraints (e.g.,

real human mobility data is considered personal data,

which is restricted by privacy laws under the Eu-

ropean Data Regulation (European Parliament and

Council, 2016)).

The human mobility models describe typical be-

havior, e.g., according to a training dataset or to pre-

Gonçalves, P. and Baier, H.

Applying Activity-Based Models to Integrate Labeled Preset Key Events in Intra-Day Human Mobility Scenarios.

DOI: 10.5220/0011850000003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 281-288

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)

281

deﬁned rules. On the other hand, when individuals

decide to alter their normal behavior, it might not be

considered in these models. For instance, from the

psychological forensics perspective, criminal behav-

ior is hard to model as criminals adapt to their envi-

ronment by changing their actions to elude prosecu-

tion, as described by Holmes and Holmes (Holmes

and Holmes, 2008). From another view, in machine

learning algorithms the outliers describe disparate

data that opposes the data from the trained models,

i.e. anomalies in a controlled environment. In the

scope of our work, we deﬁne all non-typical behavior

as the key events to be integrated in a mobility dataset.

Contemporary mobility frameworks are able to cre-

ate synthetic mobility data, but we identiﬁed a limited

support to integrate customized key events. Thus, we

propose in this work a methodology to address this

gap.

The rest of this paper is structured as followed. In

the subsequent Section 2 we discuss the current state

of the art in modeling human mobility and the asso-

ciated activity schedulers. Further, we review in Sec-

tion 2.2 on previous work done on intra-day activity

schedulers and discuss strengths and weaknesses of

available mobility frameworks regarding the injection

of preset activities and identify a research gap on inte-

grating controlled key events. Based on the review of

related work we suggest a methodology in Section 3

and thus our attempt in closing this gap. We validate

our method by providing a possible implementation

in Section 4.1 using Python 3.10 and validate against

predeﬁned requirements from Section 3.3. Finally, we

discuss our results and derive possible future work in

Section 5.

2 RELATED WORK

In this section we present related work to our ap-

proach. We screen the online databases of the four

largest IT publishers IEEE Xplore, ACM Digital Li-

brary, Elsevier Science Direct and Springer Link and

two scientiﬁc search engines Google Scholar and Re-

searchGate between years 2010 and 2022. Our re-

sult is that activity-based models are often designed

in integrated systems, hence a data scientist may pro-

vide variable initial settings, but the systems feature

limited means to inject static activities (i.e. the key

events) in the activity planning phase at the agent’s

individual level.

2.1 Modeling Human Mobility and

Activity Schedulers

One intuitive approach to model human mobility of

single individuals is the use of activity-based models

as a bottom-up approach for simulating human deci-

sions in a resource-limited environment (Castiglione

et al., 2015). According to Zheng et al. (Zheng

et al., 2013) agent-based models are often rooted on

activity-based models, where agents are autonomous

individuals that may include their past experiences in

future decisions, e.g., using reinforced learning algo-

rithms. Agent-based models are a common choice

within the research community to simulate human

mobility of small groups up to complete populations

concurring in a joint environment with limited re-

sources, e.g., temporal and spatial constraints.

The models are often implemented in integrated

systems, i.e. the system gradually adapts to changes

without user intervention. For instance, the authors

Luca et al. identify in their work (Luca et al., 2021)

missing control in solutions based on deep learning.

Thus, the systems offer limited means to induce con-

trolled activities and therefore force agents’ behaviors

to be adapted accordingly. Further, these models are

data-driven and rely on high amounts of representa-

tive data obtained from the target region matching the

desired use case.

The review on human mobility models by Sol-

maz and Turgut in 2019 (Solmaz and Turgut, 2019)

determined a need in creating scenario-speciﬁc and

realistic human mobility models. According to their

work, available intra-day mobility models mainly fo-

cus on creating typical behavior, e.g., in a sleep-

work-leisure cycle as in the Working Day Movement

Model (Ekman et al., 2008). Modeling non-typical

behavior, i.e. inducing scenario-speciﬁc activities in

available models, are currently not in focus of current

research, although Ekman et al. (Ekman et al., 2008)

determined a general research trend towards scenario-

speciﬁc models. These models might be used to gen-

erate mobility datasets simulating the usage of a wide

range of devices and vehicles, according to the current

mode of transportation.

Activity schedulers are used to plan and execute

activities from a set of possible activities, where ac-

tivities represent the coarse choices of an individual.

This can be performed on a daily basis according to

the individual’s current needs and desires. For in-

stance, if one individual’s need is to buy groceries,

then an activity scheduler would search for grocery

stores in the near vicinity according to his desires.

In most cases, humans start using different modes

of transportation to their destination, and again ac-

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

282

cording to their desires. After choosing an activity,

a scheduler would then calculate the estimated time

slot of the activity and executing it, considering any

changes in the environment, e.g. trafﬁc jams may af-

fect the travel time or alternate faster routes may be

the better choice. In most mobility frameworks, an

intra-day activity scheduler is often combined using

a microscopic mobility models to generate the track-

ing data at high spatial and temporal resolutions, e.g.

applying routing mechanisms in a vehicular road net-

work (Castiglione et al., 2015).

2.2 Agent-Based Systems

An example of an agent-based framework is SimMo-

bility created by Adnan et al. (Adnan et al., 2015) us-

ing multiple modes of transportation. This framework

is integrated and consists of different sub-models to

simulate a range of complex context-aware decisions

and mobility at different time steps in a demand-and-

supply paradigm. Changes in one of the sub-models

are propagated to the others sub-models. Due to it’s

system design, it is not intended to adapt the schedul-

ing process to integrate further preset activities.

The open-source project MATSim (Horni et al.,

2016) is an agent-based modeling framework to ef-

ﬁciently simulate a wide range of use cases of intra-

day activities, using different modes of transportation.

The scheduler employ a set of different strategies in

a optimization problem setting, even when the initial

data is not complete, as agents automatically adapt to

ﬁnd better solutions. Due to it’s design, it provides a

method to interfere with the daily activity planning,

but as induced activities might not provide an opti-

mal solution it may not be selected and is not suitable

including preset activities in the scheduling process.

Another tool is the work by B

osche et al. (B

osche

et al., 2012). Their tool creates synthetic data based

on statistical properties of origin-destination pairs on

travel surveys. While their synthetic approach resem-

bles the statistical properties of real data, it does not

provide means to inject preset data, individual behav-

iors and event-based delays, e.g., roadblocks.

The authors Drchal et al. introduced in their

work (Drchal et al., 2019) an activity-based sched-

uler applying various strategies to select and plan a set

of intra-day activities, similar to the mid-term model

of the SimMobility (Lu et al., 2015) framework. In

detail, the authors (Drchal et al., 2019) select a set

of possible next activities, calculate the estimate du-

ration and travel time. From the given set of ac-

tivities they choose the one with the highest attrac-

tiveness, according to the statistical properties of real

data. Although, the scheduler needs real data to train

the model, it can be used to generate additional data

outside the training area, supposing that the sub-area

is an adequate representative of the complete simu-

lated area.

The authors Al-Kuwari and Wolthusen describe

in their work (Al-Kuwari and Wolthusen, 2010) that

the geospatial forensic analysis of conﬁscated de-

vices rarely have a complete tracking dataset available

and forensic experts need to manually reconstruct the

tracks. Therefore, they introduce a multi-modal trace

reconstruction algorithm by applying different modes

of transportation, i.e. mobility models and recon-

structing data using a probabilistic approach. While

this approach focuses on modeling mobility between

two positions, it does not provide means to integrate

additional activities.

In our review we could not ﬁnd any method to

adapt the scheduling process for mobility activity-

based models and thus integrating preset activities in

the activity planning phase. Thus we suggest a pos-

sible approach in the following Section 3 where we

attempt to close this research gap.

3 METHODOLOGY

We suggest in the following a method to extend avail-

able mobility models to integrate preset activities in

the generation process. We contribute to close the

research gap and propose a method which integrates

a set of static key events and models additional syn-

thetic events around them. The resulting set of preset

and generated activities for each agent is then the ba-

sis for generating synthetic human mobility at a mi-

croscopic level.

In detail, we discuss the scheduling order in Sec-

tion 3.1 using a hierarchical approach. Next, we de-

scribe the scheduler using the intermediate activity

model to iteratively generate further activities in free

time slots in Section 3.2. In Section 3.3 we deﬁne

the attributes and constraints of activities needed to

be considered for the intermediate activity model. Fi-

nally, we discuss how an agent’s individual preference

might be considered in scheduling additional activi-

ties in Section 3.4

3.1 Model Activity Order

Activities describe coarse tasks, which differ in the as-

pects when, where and how they can be executed and

therefore have several constraints, i.e. how ﬂexible

they are in the scheduling process. The constraints

can be described in time (e.g., working and service

hours), space (e.g., swimming activities can only be

Applying Activity-Based Models to Integrate Labeled Preset Key Events in Intra-Day Human Mobility Scenarios

283

performed at public swimming pools, at home or at a

public lake) and type (e.g., transport activities highly

depend on the available modes of transport). A ran-

dom scheduler may produce activity plans that could

exclude vital activities, e.g., work activities can be

primary planned during working hours at predeﬁned

days of the week.

Therefore, we propose that activities should be

scheduled according to their importance and/or by

the degree of constraints, similar as in the ILUTE

Model (Salvini and Miller, 2005). To simplify the

scheduling, we use a numerical scoring to aggregate

different constraints. Thus, activities with more con-

straints are scheduled with higher priority than activ-

ities with fewer constraints. Using preset activities

as key events in activity plans are non-optional and

should always be scheduled ﬁrst, that is before adding

other activities. This concept can also be further ex-

panded to adjust to customized scheduling strategies

and adapting the scoring, e.g., prioritizing activities in

certain situations.

3.2 Intermediate Activity Model

To include preset activities, we adapt the schedul-

ing process in using higher-constrained activities and

schedule other lower-constrained activities in be-

tween, similar as in the intermediate stop location

model (Castiglione et al., 2015) and name it accord-

ingly: intermediate activity model. The intermediate

stop model is originally used to simulate additional

stops in a tour between two successive points. Thus,

additional stops (or activities) are detours where the

destination is usually a work place, school or a return

home location.

Activity Time Windows. According to Castiglione

et al. (Castiglione et al., 2015), one possibility to de-

termine when activities should be scheduled, is the

use of time windows. This time window speciﬁes

a time slot where additional activities and the corre-

sponding trips might be generated. Preset activities in

general might not start and end at ﬁxed time periods

and therefore a scheduler needs to dynamically adapt

the time windows according to the actual starting and

ending times of consecutive activities. When new ac-

tivities are added to the activity plan, the scheduler

adjusts the time window accordingly.

3.3 Attributes of Activities

Inserting additional activities between two successive

activities according to the intermediate stop model

forces a scheduler to be context-aware. This may

cover the previous scheduled activities (e.g., reason-

ing chains), the mode of transport utilized in this tour

(e.g., for mode chains modeling (Song et al., 2021)),

the potential activities that might be chosen as inter-

mediate stops in this time window (e.g., constrained

by working and service hours), estimated travel times

to reach the next destination, constraints placed on

the environment (e.g., limiting the number of agents

per area) and the agent’s proﬁle (i.e. his desires and

needs). In the following, we describe the minimal

necessary attributes for using the intermediate stop

model inside a scenario.

Each activity A consists at least of a duration d,

timestamps t at a starting point p

and ending p

point and an activity type class. Further we assume

that each activity has the following constraints and at-

tributes: Exclusivity, chain of reasoning, physical and

time constrains, labels. In the following, we deﬁne

the given constraints and attributes.

Exclusivity. At any time t, there is at most one ac-

tivity assigned and therefore we can exclude that more

than one activity is performed simultaneously, i.e. we

can deﬁne an injective function f : T 7→ P with t ∈ T

and A ⊂ P, where T is the complete set of timestamps

and P the activity plan containing all scheduled activ-

ities A. If an activity consists of a set of sub-activities,

then these sub-activities should consider the tempo-

ral and spatial boundaries of it’s parent, i.e. start-

ing and ending times and positions. With this deﬁ-

nition, we can order all available activities according

to their starting times (or ending times) and conclude

that two consecutive activities A

and A

i+1

do not con-

tain any additional activities in between or overlap-

ping each other. Based on the previous deﬁnitions, we

can now insert additional activities in the time win-

dow between two consecutive activities as depicted in

Figure 1, i.e. after the activity A

and before the activ-

ity A

i+1

Chain of Reasoning. The integration of new activi-

ties in a given time slot may induce several problems.

Any new activity A

new

, that is not in the close vicinity

of the previous activity, may induce additional activ-

ities, e.g., creating a transport activity (I and III in

Figure 1) according to the available modes of trans-

port to reach a destination. Any newly added activity

should contain sufﬁcient time for an agent to reach

the consecutive activity with any available modes of

transportation.

Physical and Time Constraints. Human mobility

in general is not without limitation. According to

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

284

i+1

new

time gap

I III

Figure 1: Intermediate tour model: A new activity A

new

can

be inserted between the given activities A

and A

i+1

having

a ﬁxed time gap. The scheduler needs to consider potential

transport activities from A

to A

new

(I) and from A

new

to A

i+1

(III). The step in II can be repeated to recursively add more

activities until either the time gap or the agent’s needs are

exhausted.

the mode of transportation or the activity, some ar-

eas or paths might not be accessible for speciﬁc time

slots. For instance, shopping centers may have time

restrictions and individuals are not permitted to enter

this area outside service hours. Further, the choice

of transport also inﬂuences variables like travel time

(e.g., limited by speed limits), route choices (e.g.,

choosing routing alternatives during a trafﬁc conges-

tion) and path choices (e.g., in general vehicles travel

on a directional network). This rule set should be ex-

tended to match the desired use case and is therefore

not complete.

Labels. Any activity contains a description (e.g., a

label) about the type of activity. This is not only

needed by an algorithm to properly recognize it’s

type, but also include some level of documentation

for the data scientist to reproduce the results. Appro-

priate descriptions might inﬂuence how, where and

when activities are planned and executed. For in-

stance, according to the chain of reasoning, some ac-

tivities are dependent on previous decisions and labels

aid to properly identify them.

3.4 Modeling Agent Preferences

Each agent has his own desires and needs, therefore

we must consider that some agents might have pref-

erences or restrictions in their choices of the activity

type and mode of transportation. It depends, accord-

ing to Castiglione et al. (Castiglione et al., 2015) on

variables like income, social statuses, individual ac-

cessibility and availability to use different modes of

transport and demographic data. This list should be

extended to ﬁt the current use case.

The agent’s preferences might inﬂuence how,

when and where activities are scheduled and thus lim-

iting the choices for a scheduler. For instance, de-

pending on the agent’s occupation, we ﬁnd different

time windows when a certain work might be sched-

uled, e.g., night-shifts. Going further, we may iden-

tify, that the workplace is always performed at known

location in contrast to maintenance works for the gen-

eral public, where workers respond to house calls and

constantly change their work place. Therefore, each

activity type class should be modiﬁed to ﬁt the agent’s

need in adding or changing the corresponding param-

eters, e.g., work hours differ for each agent, but are

ﬁxed.

4 VALIDATING THE

METHODOLOGY

Although the suggested methodology might be scaled

up to be used in a multi-agent system, it does not di-

rectly affect other agents. In addition, comparing the

resulting dataset with real data or other models is lim-

ited, as key events might describe non-typical behav-

ior. In this section we present a Python-based imple-

mentation and validate our methodology as presented

in Section 3 to include preset activities for simulating

a human mobility scenario.

4.1 Implementing the Activity

Scheduler

The outline of the implementation approach of

scheduling intra-day activity plans is depicted in Fig-

ure 2. It consists of a preparation phase, where the

system cleans and transforms the input data to be pro-

cessed in the next steps. During the next phase, the

system synthesizes the population, i.e. we deﬁne for

each agent their preferences and needs. In a daily cy-

cle, the activities are hierarchically scheduled accord-

ing to the model order as deﬁned in Section 3.1. Each

daily cycle ends in a simulation phase at the micro-

scopical level by generating the actual tracks. Regular

logical checks at different phases assure that sched-

uled activities are conform with the requirements as

deﬁned in Section 3.3. In the following we go into

more detail for each of the implementation steps.

Activity Preprocessing Phase. In this preprocess-

ing step, the scheduler identiﬁes the type of the preset

key events, the starting and ending locations and the

timestamps at these locations. This phase is mainly

associated with data cleaning and transforming user

input data and creates activities to be processed in the

next steps.

Applying Activity-Based Models to Integrate Labeled Preset Key Events in Intra-Day Human Mobility Scenarios

285

activtity

preprocessing

logical checking

synthesizing

population

preset activities

intermediate

activity model

executing

activity plan

logical checkingnext day

Figure 2: This chart describes the program ﬂow of injecting

preset activities. After preparing the preset data in a prepro-

cessing phase, the population is synthesized in a consecu-

tive step. In a daily cycle, the program schedules the preset

activities and gradually plans additional activities according

to their degree of constraints. Logical checks are executed

at different phases to minimize logical discrepancies in the

insertion of (preset) activities. In the last step, the activity

plan is executed using microsimulation models

Synthesizing Population. In this intermediary step,

we synthesize the population by creating agents, hav-

ing their own desires and preferences. Here we de-

ﬁne the needed global information like ﬁxed locations

(e.g., home position or workplace), when and how ac-

tivities are scheduled by adjusting the initial settings

for each agent. In this step, a data scientist can easily

create different proﬁles for each agent.

Adding Preset Activities. On a daily basis, all pre-

set activities are scheduled for the associated agents.

Between two consecutive preset activities, we calcu-

late the respective time window, where additional ac-

tivities might be scheduled. As long as the time win-

dow is greater than a given threshold, we can itera-

tively plan new activities. The threshold is a combina-

tion of a minimum duration of an activity and the rel-

ative travel times between the activities, as described

in the intermediate activity model.

Intermediate Activity Model. The intermediate

activity model describes how additional activities are

consecutively generated between two given activities.

We implemented this model as described in Algo-

rithm 1, where new activities are iteratively added.

Given the two consecutive activities A

and A

i+1

and

a potential new activity A

new

, a scheduler generates a

transport activity from the previously added activity

to this new activity A

new

and plans a transport

activity from the newly generated activity to the last

activity A

i+1

. The new activity is only added if both

travel times and the duration of the activity do not ex-

ceed the time window. The last transport activity is

only of importance, if we can not create additional

activities, i.e. the time window is too small for new

activities to be added or there are global settings that

do not permit creating more. In this case, we can ex-

ecute the transport activity.

← A

;

while ∆

gap

≤ α do

generateTransport(A

new

);

generateActivity(A

new

);

planTransport(A

new

i+1

);

∆

gap

← adaptTimeWindow();

← A

new

;

end

executeTransport(A

i+1

);

Algorithm 1: Inserting intermediate activities A

new

in a

given time window ∆

gap

of activities A

and A

i+1

Microsimulation Mobility Model. We used the

python OSMNX library to download and extract

the paths, points of interest and nodes of the

OpenStreetMaps (OSM) database (Open Street Map,

2023). This framework also provides useful data

cleaning methods, e.g., connecting street ends at in-

tersections or calculating travel times between nodes.

The concept of nodes are used in OSM to represent a

multitude of things, but we are only interested in those

representing connections in paths, e.g., intersections,

curved ways or stop-points and edges represent sub-

paths, e.g., parts of a road. The routing of agents on

a network is implemented as a weighted Dijkstra al-

gorithm, where simulated travel times are the edges’

weights between two nodes.

Logical Checking. To ensure that activities do not

contain logical faults, we deﬁne a global set of rules

prior generating new activities. The rule set can also

be used to check for any human error prior integrating

preset activities to the system during the preparation

phase. In our system design, we included the follow-

ing attributes from Section 3.3: ordered set, exclusiv-

ity and chain of reasoning.

As preset activities might describe non typical be-

havior, we did not include checks for preserving any

time and physical constraints and individual prefer-

ence, e.g., entering restricted areas after service hours.

However, those attributes are considered as rules in

the planning of new activities.

4.2 Validation

Therefore, to ensure that the implementation from

Section 4.1 of our proposed methodology (see Sec-

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

286

tion 3) provides the intended results, we validate the

resulting dataset of one agent against the requirements

from Section 3.3.

We provided our implementation with a set of

ﬁxed preset activities as our key events and generated

additional activities around these activities. Consider-

ing that preset activities might represent any activity

combinations of non-typical daily behavior, e.g., hav-

ing a day off from work and performing other activi-

ties, we decided to randomize the preset dataset with a

high variety of activity combinations. The methodol-

ogy was implemented on a Windows 11 Pro (Version

21H2) running on Intel Core i9-12900H with 64GB

RAM using Python 3.8.11, OSMNX library version

1.1.1.

The Preset Activity Dataset. This dataset contains

randomized activities, where the location during an

activity does not change, i.e. the agent starts and ends

the activity at the same location an no further track-

ing data is generated during his activity. Usually, the

transition between two consecutive activities is mod-

eled as a transport activity, given that the next activity

is not at or in the near vicinity. We restricted our test

data to be generated in a urban area of Munich, Ger-

many with approximately 11km

. Further we adjusted

the agent’s proﬁle, so that sleeping and at home activ-

ities are scheduled at the same location(, i.e. a physi-

cal constraint). To simulate a time constraint, we de-

ﬁned that sleeping activities can only be scheduled be-

tween 19:30 and 9:00 hours. The activities shopping

and work activities should only be scheduled between

8:00 and 20:00 hours. The simulation order for the

activities are: work as mandatory (highest scoring),

sleep and shopping as restricted (medium scoring),

walk and at home as unrestricted (lowest scoring).

Results. A sample intra-day activity plan contain-

ing preset and simulated activities is described in Ta-

ble 1, where preset activities are labeled with the ’pre-

set’ tag in the simulation order column. Other labels,

e.g., mandatory or restricted, describe the simulation

order for this particular activity. Further, we identify

that each activity complies with our requirements:

• ordered set all activities are ordered according to

their starting/ending times.

• exclusivity at all times, we identify at most one

single activity scheduled.

• chain of reasoning between two consecutive non-

transport activities, there is always a transport ac-

tivity, if not scheduled at the same location, which

is always true in our sample.

• physical constraints all at home and sleeping

activities are scheduled at the same location, i.e

(48.18286, 11.52519).

• time constraints sleeping (entry 3) is scheduled

between 22 and 9 hours and shopping (entry 13)

is scheduled between 8 and 20 hours. We want to

point out that the entry 11 is a preset activity and

therefore not bound to this time constraint.

5 CONCLUSION

In this paper we identiﬁed the research gap, that

the current state-of-the-art activity-based models on

intra-day human mobility do not contain a possibility

to adapt the scheduling process to include preset ac-

tivities with ﬁxed timestamps and locations. Having

preset activities allows data scientists to injecting key

events and thus provide a control mechanism in the

generation of activity plans, which then may be used

in a next step to generate customized geospatial data.

To close this gap, we proposed a methodology

based on activity-based models to address this prob-

lem and deﬁned the requirements for these preset ac-

tivities, i.e. ordering, exclusivity, chain of reasoning

and aware of time and physical constraints according

to the agent’s individual preferences. To validate our

methodology we created a Python-based implemen-

tation using the OSMNX framework for integrating

OpenStreetMaps. We purposed a hierarchical-based

intermediary activity model to add new activities ac-

cording to their degree of constraints, as described

in Section 3.1. As preset activities might represent

any combination of typical and non-typical behavior,

we included a randomized set of preset activities and

showed that the resulting set also complies with our

requirements in Section 4.2.

Although our proposed methodology is able to in-

tegrate predeﬁned activities in the scheduling process,

it is limited to non-transport preset activities that start

and end at the same location. For integrating prede-

ﬁned transport activities that contain different starting

and ending positions, we may need to consider the

reason behind this transport activity, i.e. what agent’s

need is met by this preset transport activity. This

might possibly trigger implicit activities needing to be

added and adapting the scheduling order, e.g., a trans-

port activity between home and work location would

implicate activities to be performed at home and at

work. As the integration of transport activities is not

trivial, we propose future work in further closing this

research gap.

Applying Activity-Based Models to Integrate Labeled Preset Key Events in Intra-Day Human Mobility Scenarios

287

Table 1: A sample test dataset containing preset and simulated activities from one agent in a 24 hour window. The activities

have physical and time constraints when and where they may be scheduled. Further, we identify that our requirements are

met in the generated activities, i.e. ordered set, exclusivity and chain of reasoning as described in Section 3.3.

Entry Sim. order Activity type Start Time End Time Start (lat, lon) End (lat, lon)

1 preset home 17:34:14 19:35:40 (48.18286, 11.52519) (48.18286, 11.52519)

2 restricted transport 19:35:40 19:35:40 (48.18286, 11.52519) (48.18286, 11.52519)

3 restricted sleep 19:35:40 02:53:24 (48.18286, 11.52519) (48.18286, 11.52519)

4 unrestricted transport 02:53:24 02:53:24 (48.18286, 11.52519) (48.18286, 11.52519)

5 unrestricted home 02:53:24 05:12:36 (48.18286, 11.52519) (48.18286, 11.52519)

6 unrestricted transport 05:12:36 05:13:59 (48.18286, 11.52519) (48.18469, 11.53327)

7 unrestricted walk 05:13:59 07:03:00 (48.18469, 11.53327) (48.18469, 11.53327)

8 unrestricted transport 07:03:00 07:08:47 (48.18469, 11.53327) (48.18418, 11.49277)

9 unrestricted walk 07:08:47 11:20:24 (48.18418, 11.49277) (48.18418, 11.49277)

10 unrestricted transport 11:20:24 11:25:34 (48.18418, 11.49277) (48.18286, 11.52519)

11 preset sleep 11:25:34 12:30:24 (48.18286, 11.52519) (48.18286, 11.52519)

12 unrestricted transport 12:30:24 12:34:36 (48.18286, 11.52519) (48.18005, 11.49265)

13 unrestricted shopping 12:34:36 13:15:54 (48.18005, 11.49265) (48.18005, 11.49265)

14 unrestricted transport 13:15:54 13:20:28 (48.18005, 11.49265) (48.18286, 11.52519)

15 preset home 13:20:28 17:16:49 (48.18286, 11.52519) (48.18286, 11.52519)

REFERENCES

Adnan, M., Pereira, F. C., Azevedo, C. M. L., Basak, K.,

Lovric, M., Raveau, S., Zhu, Y., Ferreira, J., Ze-

gras, C., and Ben-Akiva, M. (2015). Simmobility:

A multi-scale integrated agent-based simulation plat-

form. In 95

Annual Meeting of the Transportation

Research Board Forthcoming in Transportation Re-

search Record.

Al-Kuwari, S. and Wolthusen, S. (2010). Forensic track-

ing and mobility prediction in vehicular networks.

In Advances in Digital Forensics VI, pages 91–105.

Springer Berlin Heidelberg.

Aschenbruck, N., Munjal, A., and Camp, T. (2010). Trace-

based mobility modeling for multi-hop wireless net-

works. Computer Communications, 34(6):704–714.

Barbosa, H., Barthelemy, M., Ghoshal, G., James, C. R.,

Lenormand, M., Louail, T., Menezes, R., Ramasco,

J. J., Simini, F., and Tomasini, M. (2018). Human

mobility: Models and applications. Physics Reports,

734:1–74. Human mobility: Models and applications.

osche, K., Sellam, T., Pirk, H., Beier, R., Mieth, P., and

Manegold, S. (2012). Scalable generation of synthetic

GPS traces with real-life data characteristics. In Tech-

nology Conference on Performance Evaluation and

Benchmarking, pages 140–155. Springer.

Castiglione, J., Bradley, M., and Gliebe, J. (2015). Activity-

Based Travel Demand Models: A Primer. The Na-

tional Academies Press, Washington, DC.

Drchal, J.,

Certick

y, M., and Jakob, M. (2019). Data-driven

activity scheduler for agent-based mobility models.

Transportation Research Part C: Emerging Technolo-

gies, 98:370–390.

Ekman, F., Ker

anen, A., Karvo, J., and Ott, J. (2008). Work-

ing Day Movement Model. In Proceedings of the

ACM SIGMOBILE workshop on Mobility models,

pages 33–40.

European Parliament and Council (2016). Regulation (eu)

2016/679. Ofﬁcial Journal of the European Union,

vol. 59(L 119):1–88.

Holmes, R. and Holmes, S. (2008). Proﬁling Violent

Crimes: An Investigative Tool. SAGE Publications.

Horni, A., Nagel, K., and Axhausen, K., editors (2016).

Multi-Agent Transport Simulation MATSim. Ubiquity

Press, London.

Lu, Y., Adnan, M., Basak, K., Pereira, F. C., Carrion, C.,

Saber, V. H., Loganathan, H., and Ben-Akiva, M. E.

(2015). Simmobility mid-term simulator: A state

of the art integrated agent based demand and supply

model. In 94

Annual Meeting of the Transportation

Research Board, Washington, DC.

Luca, M., Barlacchi, G., Lepri, B., and Pappalardo, L.

(2021). A survey on deep learning for human mo-

bility. ACM Computing Surveys, 55(1):1–44.

Open Street Map (2023). Open street maps. https://www.

openstreetmap.org/. Accessed: 2023-02-10.

Salvini, P. and Miller, E. J. (2005). ILUTE: An operational

prototype of a comprehensive microsimulation model

of urban systems. Networks and Spatial Economics,

5(2):217–234.

Solmaz, G. and Turgut, D. (2019). A survey of human mo-

bility models. IEEE Access, 7:125711–125731.

Song, Y., Li, D., Cao, Q., Yang, M., and Ren, G. (2021).

The whole day path planning problem incorporating

mode chains modeling in the era of mobility as a

service. Transportation Research Part C: Emerging

Technologies, 132:103360.

Zheng, H., Son, Y.-J., Chiu, Y.-C., Head, L., Feng, Y., Xi,

H., Kim, S., and Hickman, M. (2013). A Primer for

Agent-Based Simulation and Modeling in Transporta-

tion Applications. Technical report, Federal Highway

Administration.

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

288