Accepted to International Journal of Geographic Information Science (Sept 2015)

Analysis of Human Mobility Patterns from GPS Trajectories and Contextual

Information

Katarzyna Siła-Nowicka1,4, Jan Vandrol2, Taylor Oshan3, Jed Long1, Urška Demšar1, A.

Stewart Fotheringham3

1
School of Geography & Geosciences, University of St Andrews, Scotland, UK
2
School of Energy, Environmental Technology and Agrifood, Cranfield University, UK
3
GeoDa Centre, School of Geographical Sciences and Urban Planning, Arizona State

University, Tempe, Arizona, USA

4
Urban Big Data Centre, University of Glasgow, Scotland, UK

Abstract:

Human mobility is important for understanding the evolution of size and structure of urban

areas, the spatial distribution of facilities, and the provision of transportation services. Until

recently, exploring human mobility in detail was challenging because data collection methods

consisted of cumbersome manual travel surveys, space-time diaries or interviews. The

development of location-aware sensors has significantly altered the possibilities for acquiring

detailed data on human movements. While this has spurred many methodological

developments in identifying human movement patterns, many of these methods operate

solely from the analytical perspective and ignore the environmental context within which the

movement takes place. In this paper we attempt to widen this view and present an integrated

approach to the analysis of human mobility using a combination of volunteered GPS

trajectories and contextual spatial information. We propose a new framework for the

identification of dynamic (travel modes) and static (significant places) behaviour using

1/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

trajectory segmentation, data mining and spatio-temporal analysis. We are interested in

examining if and how travel modes depend on the residential location, age or gender of the

tracked individuals. Further, we explore theorised “third places”, which are spaces beyond

main locations (home/work) where individuals spend time to socialise. Can these places be

identified from GPS traces? We evaluate our framework using a collection of trajectories

from 205 volunteers linked to contextual spatial information on the types of places visited

and the transport routes they use. The result of this study is a contextually enriched data set

that supports new possibilities for modelling human movement behaviour.

Keywords: Movement analysis, trajectories, trajectory segmentation, travel mode

classification, significant places.

1. Introduction

Understanding human mobility is a major contributor to the development of knowledge on

important issues such as the form and function of urban areas, the location of facilities and

the demand for transportation services. Human mobility has traditionally been explored by

manual travel surveys, space-time diaries or interviews (Palmer et al. 2013), all of which are

time consuming, expensive and do not provide a sufficient spatio-temporal resolution to

allow the investigation of mobility patterns in detail. The development of sensors that capture

movement information in real time and at detailed spatial and temporal scales (e.g. GPS

trackers) has changed our ability to collect movement data (Kwan and Neutens 2014).

However, the developments in movement data collection technologies are much further ahead

than current methods for extracting meaningful patterns from such data (Laube et al 2007,

Long and Nelson 2013). While recently there have been several new methodological

developments in trajectory data mining on identifying patterns and behaviours in movement

2/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

data, many of these ignore the embedding of movement into the geographical context (Purves

et al. 2014). There is a need to investigate new ways to identify movement patterns

considering movement data not only on their own, but within their environmental context.

In this paper we approach this problem through an integrated analysis of human movement

from GPS trajectories linked to contextual information. Purves et al. (2014) list several

alternative definitions of context for movement data: 1) context identified from additional

data collected simultaneously with GPS trajectories, 2) context provided by the description of

space within which movement occurs and 3) context provided through knowledge about

physical/biological properties of the movement process. In this study we define context as

option 2); that is, the contextual data used in this study provide a description of the spaces

that people move through.

A particular characteristic of human movement is the variety of different travel modes via

which movement can take place. This is of particular interest in areas such as urban planning

(Schwanen and Moktharian 2005) because residential location-choice and travel-choice may

be interconnected. It is known that the built environment and the socio-demographic

characteristics of individuals have impacts on the choice of transportation mode for daily

travel such as commuting (Ewing and Cervero 2010). This impact is an extensively

researched topic in transportation geography and has traditionally been explored using data

from surveys and questionnaires (Van Vugt et al. 1996, Rodriguez and Joo 2004, Schwanen

and Moktharian 2005, Wener and Evans 2007). New research has involved using GPS

trajectories to identify transportation modes from raw data and annotate the original

trajectories with these modes to create semantic trajectories (Yan et al. 2013). Such semantic

trajectories are used for many technical purposes: personalisation of web services (Zheng et

3/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

al. 2010), navigation (Sester et al. 2012), privacy protection (Parent et al. 2013) and inferring

social structure from trajectory proximity (Xiao et al. 2014). There has, however, been little

or no connection between these two points of view. We offer an alternative where we identify

transportation mode from real GPS data while at the same time exploring if and how this

mode is affected by the residential location of our participants.

Patterns of human mobility are linked to how people influence their spatial context and how

the spatial context influences them in return (Palmer et al. 2013). A location becomes a place

through perceptions of it and the activities of the people who use it, which in turn affects the

behaviour of the people there. Thus, human movement is not only embedded into

geographical space, but can be considered as movement between places imbued with

meaning through human activities (Gieryn 2000). From the sociological perspective, each

individual moves through a set of hierarchically ordered places that have a particular meaning

for him/her (Oldenburg 1989). The most important place, termed the first place, is home,

where an individual spends most of his/her time. The second place is work or school or a

place where a major regular activity takes place. In the process of building social capital,

each individual also frequents so-called “third places”. These places are neither home nor

work-related and are places where people spend their leisure time and socialise with others.

They can include many different types of locations: shops, cafes, bars, libraries, bookshops

(Oldenburg 1989, Holm 2013) and have become a popular subject of study in sociology,

urban geography and retail geography (Holm 2013, Laing and Royle 2013, Lin 2012,

Steinkuehler and Williams 2006). However, to our knowledge, there have been no studies

investigating third places using empirical tracking data. There are a number of technological

studies that identify the so-called significant places from movement data (places where

individuals spend most of their time and which have the highest re-occurrence rate, Liao et al.

4/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

2007), but these are not necessarily ‘third places’ and the studies are generally devoid of

context. For example, Bhattacharya et al. (2012) only use the geometrical properties of

movement inferred from GPS trajectories to identify significant places and purposely exclude

the geographical context (“instead of considering additional resources such as geometry of

Point Of Interests (POIs) or map matching techniques, to identify a place as significant, we

propose a technique that automatically extracts all significant places and their

corresponding durations” solely from users’ GPS trajectories (Bhattacharya et al. 2012,

p.399)). Umair et al. (2014) identify significant places from GPS data based on the spatial

density of GPS points in the neighbourhood of each location, again disregarding any

contextual information.

We propose to investigate movement from transportation and sociological points-of-view

using a novel framework to integrate trajectory data with contextual information. As Kwan

(2013) states, “people’s spatio-temporal experiences are influenced not only by where they

live, but also by other places they visit, when they visit these places, how much time they

spend there, what they experience as they travel between these places”. To address these

different aspects of movement, we offer a new integrated approach that involves several

perspectives.

From the technological perspective we propose a new framework for identifying and

analysing dynamic (movement) and static (places) behaviour from trajectories and associated

contextual information. We do this through a combination of trajectory segmentation,

classification and spatio-temporal analysis using movement trajectories linked to contextual

spatial data. From the perspective of transportation geography we are interested in whether

and how identified travel modes depend on the residential location of the persons being

5/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

tracked. From a sociological perspective we explore the existence of “third places”. We are

interested in the temporal dynamics and the spatial distribution of these places – as far as we

know this is the first attempt to use real movement data for this purpose. Further, we are

interested in how gender and age might affect movement patterns and report our results

segmented by these two characteristics. We evaluate our framework on a data set of GPS

trajectories from 205 volunteers in three Scottish towns who were continuously tracked in

their daily movements for a week. We enrich these trajectories with contextual information

from external sources to identify mobility patterns in the volunteers’ daily lives.

2. Related Work

The framework we employ consists of a combination of several analytical methods for

movement data: trajectory segmentation of raw data; identification of significant places; and

classification of behaviour (static behaviour - types of places and dynamic behaviour - travel

modes). In this section we review related work on these methods.

Segmentation is the process of dividing a movement trajectory into sub-trajectories called

segments, where each segment fulfils certain criteria (Buchin et al. 2011). Segmentation is

frequently derived using “movement parameters”, i.e. statistical properties of the movement

process at each trajectory point. These include velocity, speed, heading, acceleration, turning

angle, angular range, displacement, straightness index, sinuosity, tortuosity, and other locally

calculated features (Dodge et al. 2009, Dodge et al. 2012).

In transportation, a combination of segmentation and data mining is frequently used for the

classification of travel mode. Sester et al. (2012) identify important places to divide

trajectories into segments of constant travel mode. Torrens et al. (2011) use machine learning

for identification of pedestrian behaviours. Hu et al. (2004) use neural networks (Self-

6/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Organising Maps) to classify different motion types. Zheng et al. (2010) use a change-point

segmentation method to classify GPS trajectories into travel modes.

Significant places (Liao et al. 2007) can be identified from GPS trajectories using clustering

or machine learning (Ye et al. 2009, Rodrigues et al. 2014, Umair et al. 2014). Studies

identifying significant places can be broadly categorised into two groups: those identifying

personally meaningful locations for individuals (Ashbrook and Starner 2003, Kang et al.

2004, Zhou et al. 2007, Bhattacharya et al. 2012) and those finding significant places for

multi-users that identify general places of interest (Ashbrook and Starner 2002, Agamennoni

et al. 2009, Zheng et al. 2009 and Yin et al. 2014). Some studies use check-in histories (Lian

and Xie 2011) or user similarity (Shaw et al. 2013) to identify attractive locations. Andrienko

et al. (2013) propose a visual analytics methodology for identifying significant places.

Important places have also been identified from mobile phone trajectories (Isaacsman et al.

2011).

It is also important to categorise the activity that occurred in each of these places. Zhou et al.

(2007) and Huang et al. (2013) classify activity places into major and minor places using the

time spent in each place and the frequency of an individual re-visiting a place. Liao et al.

(2007), Kang et al. (2004) and Zhou et al. (2007) assign the categories of ‘home’ or ‘work’ to

an activity place using ‘dwell-time’ as the main class separator. Our proposed framework

incorporates each of these steps: segmentation, travel mode classification and the

identification and categorisation of significant places.

3. Data

7/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

To evaluate our proposed framework, we used data collected in a GPS travel survey and

external data sources describing the environment within which the movement occurred.

3.1. GPS travel survey

In 2013 we gathered mobility data from 205 participants living in the three largest towns in

the Kingdom (county) of Fife in Scotland: Dunfermline, Glenrothes, and Kirkcaldy. These

towns were selected due to their different socio-economic characteristics which we expected

would be reflected in the movement behaviour of their residents. Dunfermline (pop. 49,706)

is an old town located about 20km north of Edinburgh. A large proportion of the inhabitants

commute to Edinburgh for work, either by car or public transportation. Glenrothes (pop.

38,679) was established in the late 1940s as one of Scotland’s post WWII new towns and is

an industrial centre approximately half way (approx. 50km) between Edinburgh to the south

and Dundee to the north. Kirkcaldy (pop. 49,709) is an old town located 30km north east of

Edinburgh, across the Firth of Forth. It has a small medieval centre surrounded by 17th-19th

century developments and large areas of modern housing and industrial estates.

We designed our GPS survey to recruit volunteers of all ages and demographics. Many

contemporary movement studies use trajectories collected from social media (Twitter,

Foursquare), which creates an age bias towards the younger population (Bricka et al. 2012).

In order to counteract this bias, we decided to recruit participants using a traditional method:

by sending invitations via the mail to a randomly selected sample of participants. Further, we

wanted our sampling to represent the spatial distribution of inhabitants in the three towns so

that invitations were sent to a number of randomly selected addresses obtained from the

publically available Electoral Register of Scotland which were spatially distributed to reflect

population density within data zones and postcodes in each town. Data zones represent the

8/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

smallest census unit in UK, while postcodes are even smaller spatial areas, widely used for

geocoding, but they are not a perfect subdivision of data zones.

A maximum response rate for a traditional household or travel survey in UK is held to be

around 60% (Anderson et al. 2009, Dunstan 2012). Out of these responders, at most 20%

could be expected to be willing to carry a GPS device and/or perform the travel survey

(Bonnel et al. 2009). This means that we could expect a maximum response rate of 12%; in

the end, however, we achieved only a 4% response rate. We sent 6,000 invitations, equally

divided between the three towns. Out of this, 252 (4.2%) volunteers responded positively and

upon being sent GPS trackers, 205 (3.4%) trackers were returned with usable data. The low

response rate probably reflects both the perceived technical nature of the data-gathering

process and the intrusion into people’s privacy.

In order to maximise the response rate and capture the widest possible audience, we did not

collect any personal information from the respondent, nor did we request a travel diary to be

completed. We expected to be able to compensate for the lack of extra information by using

open data from governmental sources, in particular the Electoral Register of Scotland. We

were able to obtain age information from this register although for some participants this

information was not publically available. Table 1 shows the breakdown of participants by

gender and location.

Table 1 somewhere here.

The participants were asked to carry GPS trackers continuously for a period of seven

consecutive days in October or November 2013. We used i-Blue 747 ProS GPS loggers

9/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

which are equipped with motion sensors and were programmed to record the GPS location of

the individual every 5 seconds unless the device was in a stationary position for more than 2

minutes. Trajectories from all 205 participants in the main experiment comprise 3,869,831

raw GPS locations, where each location record consists of participant ID, latitude, longitude,

elevation, date and time. Figure 1 shows the spatial extent of collected trajectories.

Figure 1 somewhere here

Prior to the main data collection we also performed a pilot experiment, where volunteers

were tracked for a week. These pilot data1 were used to familiarise ourselves with the

operation of the trackers and were used as test data during the development of our

framework. The second aim of the pilot experiment was to determine the temporal sampling

rate: as per literature on GPS surveys we tested sampling rates of 1s (Krygsman and Nel

2009; Rasmussen et al. 2013), 5s (BMCT 2012), 10s (Marchal et al. 2008) and 30s (Itsubo

and Hato 2006). At the rate of 1s and tracking duration of one week, the pilot participants

collected on average 120000 GPS points, which exceeded the storage capacity of the tracker

(8Mb). A 5s sampling rate produced on average 19000 data points and filled 30% of the

storage capacity. Longer sampling rates (10s, 30s) produced data that were not accurate

enough to separate movement modes. As in similar studies (Bohte and Maat 2009), the

shortest sampling rate (1s) used up battery at a very fast rate. Based on these results from the

pilot experiment, we chose the 5s sampling rate for the main study, which had enough

accuracy for our purposes, did not drain the battery too quickly and did not exceed the

storage capacity of the tracker.

1
To support open source trajectory analysis we plan to make tracking data from the pilot participants freely available (with their
permission) at the Crawdad platform (www.crawdad.org, Kotz et al. 2004) on publication of this paper.

10/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

3.2. External Data

We expected that many participants would use public transport. Because of this, we obtained

the National Public Transport Access Node (NaPTAN) data (which contain locations of all

stations and stops of public transport) as an external source.

To be able to classify the types of significant places of individual participants, we used the

Point Of Interest data set for Scotland, produced by the Ordnance Survey UK (OS). These

data contain information about the locations of Points Of Interest, a hierarchical classification

of these (9 groups, 52 categories and 616 classes), the positional accuracy of the points and

the road network reference of the locations of each Point of Interest.

We augmented the OS Point Of Interest data by generating a supplementary Places Of

Interest data set through visual exploration of Google Maps and Openstreetmap. Note that

many Places of Interest are of an irregular size and shape (e.g. parking spaces and shopping

centres) and because of this they are better represented as polygons rather than points.

Matching a movement stop in the form of a segment of a GPS trajectory to a polygon

produces fewer errors than matching the segment to a point (fig. 2). We therefore created our

Places Of Interest as polygons rather than. We further digitised the OS Points Of Interest into

respective polygons and merged the two data sets. In the rest of the paper we refer to the

combined data of Points and Places of Interest as POI data.

Figure 2 somewhere here

11/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

The POI data consists of a set of different types of public spaces, such as shopping centres,

grocery shops, leisure centres, churches, hospitals, schools, etc. We grouped places into four

activity types as per table 2.

Table 2 somewhere here

4. Framework

The framework consists of three phases (fig. 3). The algorithms used in the framework are

provided as pseudocode in the Supplementary Online Material.

Figure 3 somewhere here

4.1. Phase One: Separating Dynamic and Static Behaviour

We consider the movement of one participant across the entire survey period as one

trajectory, which is partitioned into homogenous segments that correspond to different

movement modes. The partitioning process scans through trajectory points and when it

identifies a change in the spatio-temporal distribution of trajectory points, it creates a new

segment. For this, we define a new measure of the density of the logged positions in the

neighbourhood of each trajectory point, the Spatio-Temporal Kernel Window (STKW)

statistic. To calculate STKW values, we order all trajectory points by time. Then, for every

point we search in both directions along the trajectory and count the number of points within

a specified threshold distance from the original point, A, in figure 4a. Our threshold distance

was set to 25m, which is sufficient to distinguish stops from slower (walk/run) and faster

movement modes (vehicle transport) considering the chosen sampling frequency (5Hz). As

soon as a point further away than the specified distance is encountered, the count in that

direction stops. This differentiates our method from a point buffer which would take into

12/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

account points in the nearest neighbourhood that belong to other visits (fig. 4b). Figure 4c

shows changes in STKW for an example trajectory: these are very sudden when travel mode

changes, which allows us to identify breakpoints between segments.

Figure 4 somewhere here

To determine the start and end points of segments, the algorithm looks for maximal changes

in STKW values. It scans each trajectory using two moving windows - one facing backwards

and one forwards and then sums the STKW values within both windows. By comparing the

two sums, the algorithm decides if the point can be classified as a breakpoint between two

different travel modes. If the density of points (given by the STKW sum) on one side of the

current point is much higher than on the other side, then a point is designated as a breakpoint.

Sometimes the change in travel mode is more gradual and STKW builds up over several

points. To counter this, the algorithm searches through subsequent points for a point with the

greatest difference between the left and right STKW totals to become a new breakpoint.

Data collection was frequently temporarily halted and resumed at a later point in time from a

different location. This occurred for many reasons, including cold starts (i.e. a GPS tracker

needing to fix its current position before starting to collect data), movement inside buildings,

trackers running out of battery or being turned off by participants. Because of this, additional

breakpoints had to be introduced to split the segments into parts describing continuous

movement without temporal breaks. For this, when two temporally consecutive points within

a movement segment were located more than 280m apart from each other (maximum

distance that an object moving with 200km/h can cover within 5 seconds, which was the

sampling frequency of the trackers), the segment in question was split into two. Successive

13/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

breakpoints were ignored, thus eliminating outlying segments resulting from fake movements

when the tracker was taken indoors.

Finally, we classified segments into three movement mode categories (vehicle movement,

walk/running and a stop) using two movement parameters (Dodge et al. 2009): the median

speed and the median distance from the geometric centre of the segment. The tracker,

however, can still suggest movement when the person is stationary because of the nature of

GPS systems (e.g., multi-path effects, urban canyons, terrain obstructions). These errors can

produce high and low outliers and affect the calculation of the average speed. The median

speed therefore better represents the predominant speed throughout the segment. For this

reason, the second movement parameter, relevant for the identification of stops, is the median

distance of trajectory points in the segment from the geometric centre of the segment. Since

stop segments have their points clustered around one particular location, the median distance

from the geometric centre is able to separate them from segments representing movement.

Median distance is also less sensitive to outliers than are average distance or standard

deviation.

We used these two movement parameters to train a feed-forward neural network with back

propagation as a learning method (Haykin 2008). Neural networks are frequently used for the

classification of trajectory data due to their ability to deal with missing data and outliers

(Karlaftis and Vlahogianni 2011). Our neural network was composed of an input layer with

two neurons for each of the two movement parameters a hidden layer with three neurons and

an output layer with as many neurons as there were categories for the segment. We tested

several different configurations of the hidden layer, but found that the layer with three

neurons performed best. Higher numbers of hidden neurons led to over-fitting the network.

14/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Lower numbers of hidden neurons created a too generalised network which was unable to

classify border cases correctly. Initially we used one network to classify all types of classes

(movement classes and stop segments) and compared the results with the actual class of

objects in the training data set. This resulted in 13% misclassified objects. We therefore re-

designed the network to consist of two separate networks, one to separate stops from

movement and a second one to classify movement segments into walk/vehicle classes.

Resulting errors were 4% for stop/movement classification and 1% for walk/vehicle

classification. More details on this algorithm are provided in the Supplementary Online

Material.

The training set contained 250 manually classified segments and had an equal distribution of

classes. Once the network was trained, we used it to classify the remainder of the data set

comprising of 16789 individual segments. The results of this segmentation served as input in

the two consecutive phases, the analysis of places and the analysis of movement as shown in

figure 4.

4.2. Phase Two: Analysis of Places

In the next step we identified the locations of significant places and categorise them

according to their importance by using external spatial data (fig. 5). Note that places are not

points, but are represented as stop segments, i.e. sub-trajectories.

Figure 5 somewhere here

To identify significant places, we calculate the frequency of re-occurrence and the amount of

time spent in a location for all the stop segments from phase 1 and use these two measures as

15/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

a proxy for significance. The most frequently visited place with the longest total duration of

visits is considered to be home. Home locations derived in this way were compared to

addresses of participants in order to assess accuracy of our framework. We report on

accuracy in section 5.2. Theoretically, some individuals might spend more time in their place

of work or major daily activity (SP1) than at home. However, given the very high accuracy of

home identification (see section 5.2.) the potential for confusion between home and SP1 was

very small. An alternative possibility for home identification would be to identify places

where participants spend the night. However, as we base our classification on the sociological

theory of places (Oldenburg, 1989), we adopted the definition of home as the place where a

person spends most of his/her time. This also resolves theoretically possible issues of

participants who may spend nights at work and those who may live away from their partners

or families and spend certain nights at home and others at partner’s or family homes (e.g.

weekly commuters).

Home locations were excluded from further analysis, leaving us with a set of places that were

categorised using external contextual information. This process was based on the automatic

labelling of stop segments with information from our POI data set, which consisted of

polygons corresponding to interesting locations. All stops (sub-trajectories) found within the

proximity (50m buffer) of polygons representing the POIs (fig. 2) were assigned the activity

label based on the type of the POI (table 2).

The remaining uncategorised stops were compared to the NaPTAN database, described

above. If a stop was located in close proximity to a public transport stop, it was classified as

representing the corresponding mode of public transportation (waiting for bus/train). The

vehicle segment following two consecutive public transport stops was also renamed as

16/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

bus/train travel. The remaining stops were investigated for traffic patterns. In vehicle travel,

the movement is often interrupted with shorter stops, such as waiting at traffic lights or at the

entrance to a roundabout. We identified such traffic stops by selecting all stop segments that

were of less than 2 min duration and that occurred either between two segments previously

classified as vehicle movement or between a vehicle movement segment and a bus stop

(identified in the previous step). Further, if the transportation mode of the previous segment

were found to be a bus, the current stop and the following vehicle segments were reclassified

as bus travel. Stops that could not be identified through this procedure remained as

unidentified stops.

In order to investigate the existence of the “third places”, we used all stop segments (except

home) to identify the first three significant places (SP1, SP2 and SP3) for each participant.

These we defined as the three most frequently visited places with visits of longest duration

after home, while excluding places unsuitable for intentional socialising. This means that if a

place that fitted the definition of a significant place in terms of frequency and duration of

visits was a non-socialising place (e.g. a bus stop, a train station, a grocery store), it was

excluded from the set of significant places. Further, we limit ourselves to only three

significant places based on the longest duration of visits. These durations were on average of

212.6/48.5/18.4 min for SP1/SP2/SP3 on weekdays and 165.3/66.7/30.1 min for

SP1/SP2/SP3 on weekends. As SP4 had less than 5 min average duration for both weekends

and weekdays, it was considered too short to be included.

We expected that SP1 would be the location of either work or school (i.e. the location of the

main daily activity). However, since our participant sampling was voluntary, our

demographics included a disproportionate number of elderly people, for whom we were not

17/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

able to fully ascertain a matching between SP1 and work. Many such participants had leisure

places or shops as their SP1, which probably suggests respondents who are retired or have a

non-traditional working arrangement (e.g. working from home). We had no way of separating

these participants from those that worked in a particular location outside home (their SP1).

For this reason, we discuss SP1 together with SP2 and SP3.

4.3. Phase Three: Analysis of Movement

In the final phase we categorised the mode of transport on movement segments into

walking/running and vehicular movement, further subdivided into public transport and traffic

(fig. 3). The walking/running segments were already identified in the travel mode

classification. The vehicular movement was based on travel mode classification and the

identification of public transport and traffic stops. In the final step, consecutive segments of

the same category were merged and their attributes recalculated to reflect the newly formed

longer segments.

5. Results

We report results for the two steps: analysis of movement (travel mode classification) and

analysis of places (identification of stops).

5.1. Analysis of movement – travel mode classification

To prevent the identification of individuals or groups of individuals, we show all temporal

results as average times per day, broken down per weekdays and weekends as well as per

gender, and residential location. Table 3 shows the average time per day spent in each travel

mode. In general, there is more vehicle movement (both driving and public transport) during

the weekends than during the week, while walking averages are approximately the same. Of

18/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

note is a large increase in average time spent on public transport during the weekends

compared to weekdays for participants from Dunfermline which is in contrast to expectations

that public transport would be used primarily for commuting to work. We could not identify

any particular differences in the average use of travel modes between men and women.

Table 3 somewhere here

5.2. Analysis of places – identification and classification of stops

In order to estimate the accuracy of our place identification procedure, we compared the

centroids of home stops to the actual locations of the homes obtained from Openstreetmap

(OSM). Table 4 shows the results at spatial scales of 50, 100 and 200m with the highest

percentage of correct identification at 200m (91.21%), which demonstrates that we correctly

identify home locations within 200m in over 90% of the participants. This is comparable or

higher than similar studies: for example, Bohte and Maat (2009) report a 74% accuracy of

identifying homes from the GPS trajectories of 1104 people by extracting homes as ends of

trips. Liao et al. (2007) identify significant places with 90% accuracy although their sample

size was very small and included GPS trajectories of only four participants and their accuracy

assessment was undertaken for labelling (e.g. home, work, etc.) rather than for the locations

of places. Upon visually inspecting locations of some of the misclassified homes, we

speculate that the 90% accuracy was not an artefact of our data, but potentially related the

choice of OSM as the “ground truth” data. Geocoding street addresses is done in OSM by a

combination of actual points and interpolation, which leads to errors, as OSM automatically

excludes buildings of a certain size or type from its geocoding procedure (see Barron et al.

(2014) for known problems in OSM geocoding).

19/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Table 4 somewhere here.

We calculated the average time per day spent at home per participant (fig. 6). Most

participants spent a reasonable amount of time at home (mostly between 10-16hrs) and most

people, as expected, stayed at home longer during the weekends than during weekdays (with

some exceptions, e.g. young females from Kirkcaldy).

Figure 6 somewhere here.

We also investigated the average times per day spent in SP1, SP2 and SP3, broken down by

weekdays and weekends (fig. 7). Interestingly, the participants spent a relatively large

amount of time in their SP1, regardless of the weekday/weekend split. SPs however are

individual and fixed for each person, that is, each participant has the same three SPs

regardless of the weekday/weekend break down. Since we are taking averages, it is not

possible to say if the same participants that spend on average a lot of time in their SP1 during

the week are the same participants that spend a lot of time there during the weekend.

Participants tend to spend more time in SPs 2 and 3 on weekends than on weekdays. Note

also women aged 60-69 from Glenrothes: their averages for SP1 and SP2 are similar for

weekdays and they spend a lot of time in their SP1 weekends, which might suggest that we

captured the pattern of retired people who do not go to work but spend their time in the same

locations regardless of the day of the week.

Figure 7 somewhere here

20/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

We were further interested in the spatial distribution of SP1, SP2 and SP3. Figures 10 and 11

show these distributions as kernel density estimates (KDE) for males and female participants

respectively. We used KDEs in order to prevent identification of individual participants. The

KDE maps have the cell size of 900m and use a 2000m radius in order to mask the exact

locations of individual travel while still providing a picture of the hotspots of the SP

distribution.

Figures 8 and 9 somewhere here

For Dunfermline we expected the SPs to be both within the town of Dunfermline and in the

city of Edinburgh. The first columns in figures 8 and 9 confirm this hypothesis for both men

and women. However, men have a more widespread spatial distribution of SPs in Edinburgh,

while women’s SP hotspots are limited to the centre. We speculate that this could be a

consequence of using public transport rather than driving. Most trains coming into Edinburgh

from the north via Fife only stop at the two main train stations in the centre of Edinburgh

(Haymarket and Waverley). Any participants taking these trains would therefore likely have

their SPs in the vicinity of these two stations. Of note also is a small cluster of men’s SP

hotspots in and around Glasgow (in the lower left corner of the maps in the first column of

fig. 9). These places are present through all levels, from SP1 to SP3 and most of them belong

to the same few participants. This is a common pattern: if one individual has a SP1 in a

certain location, it is likely that his/hers SP2 and SP3 will be in or near that same location.

Glenrothes participants were expected to have their SPs within the town as well as externally

around Fife. This is confirmed by patterns for both men and women (second column in figs. 8

and 9). The main group of SPs is in Glenrothes, but there are smaller hotspots in and around

21/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Edinburgh and Fife towns (Kirkcaldy, Cupar and Dunfermline). There are also some

participants with SPs in Stirling (the hotspot near the westernmost point of the Firth of Forth)

and for men if SP1 is there (possibly indicating a place of work), SP2 and SP3 are also there.

For women, Stirling lacks a hotspot in SP1 distribution, suggesting that some female

participants travel to Stirling for leisure but do not work there.

Because of a lack of competing alternatives, Kirkcaldy residents were expected to have their

significant places mostly within the town. This is true to some extent for men and even more

for women (third columns in figs. 8 and 9).

Figure 10 somewhere here

To investigate the differences in spatial movement patterns further, we calculated the average

distance from home to SP1-2-3 respectively. Figure 10 shows radar plots of the average

distance broken down by residential location, gender and age. Home is in the centre of each

of these plots and the coloured lines show average distances for the three SPs for each age

group. Distances are up to a maximum of 50km from home. Given the total time spent at

home and in SP1, SP2 and SP3 we can consider a set of these four places together as a proxy

of an individual’s activity space. An activity space is a set of all areas within which an

individual has direct contact with others during his/her daily activities (Golledge and Stimson

1997). Considering our definition of SPs, it is likely that the first three together with home

approximate an activity space fairly well and this allows us to explore how these activity

spaces range in geographical size by age, location and gender. The results indicate some

interesting facets of different activity patterns by gender, age and location.

22/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

In Kirkcaldy the activity spaces of females tend to be more circumscribed than those of males

whereas the opposite is the case in Glenrothes. In Dunfermline, the activity spaces of the two

genders are roughly the same in extent but vary in terms of the distances travelled to the

various SPs. Females travel further to their most common social destination than do males

but males travel much further to their second most common destination than do females. This

could either be due to females facing a greater scarcity of local employment and therefore

commuting to Edinburgh or fewer females being in the workforce and their most frequent

destination being Edinburgh for social trips. Without further contextual information it is

impossible to decide on the relative weights of these two explanations although the former

seems more likely given the frequency of the trips which suggests a daily pattern of travel.

The spatial extent of activity spaces also decreases with age, as in most cases these spaces are

very local for participants of ages 60-69. In Dunfermline and Kirkcaldy the most commonly

visited places for this age group are within a kilometre from home, while in Glenrothes they

are up to 2km away. This possibly reflects that retired participants socialise in their nearest

neighbourhood rather than travelling further away. The largest activity spaces are however

not limited to a particular age group or town. The three largest spaces belong to men of age

40-49 in Dunfermline (the distances are possibly related to SPs in Glasgow (fig. 8)), women

of age 20-29 in Glenrothes (distances relating to SPs in Stirling and Dundee (fig. 9)) and men

of age 50-59 in Kirkcaldy (distances relating to Stirling and Perth (fig. 8)). Given the self-

selecting sample of respondents and lack of contextual information it is not possible to

suggest any particular age-related interpretations of these spaces except for the likely

shrinking of activity spaces linked to retirement and old age.

6. Conclusions and Discussion

23/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

In this paper we propose a framework combining movement analysis methods: segmentation,

analysing places, and identification of human mobility patterns (travel mode and places) from

a combination of volunteered GPS trajectories and contextual information. This work extends

previous studies by demonstrating a process for incorporating both movement-based and

place-based analysis into a single study to uncover new information about human movement

behaviour. We highlight how GPS tracking data can be utilised in conjunction with

contextual spatial information to study and understand individual travel behaviour and places

of interest. We highlight differences in behaviour based on residential location, gender and

age. While there are no particular differences in the use of travel mode between men and

women, our results suggest that the activity spaces of men are larger than those of women

and are larger for younger versus older adults.

This study also contributes to the literature documenting the “third places”. We develop a

methodology for capturing third places through the identification of typologically relevant

significant places (SPs) ordered by visit duration. As far as we know, this is the first attempt

to look for potential “third places” in GPS tracking data and represents a starting point for

future research into this area.

Although we highlight the potential of GPS traces for the identification of human mobility

patterns, the process is clearly not without difficulties. For instance, there are two types of

information that we decided not to include in our data collection in order to maximise the

response rate: demographic data and travel diaries. We expected that we would be able to

compensate for the lack of demographic information by accessing publicly available

governmental open data. However, since people can control the level of their personal

information that is made public, our expectations were not met. On the other hand, our

24/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

response rate was very low even without requesting such information and the balance has to

be struck between sample size and the amount of information requested from each volunteer.

An advantage of our data collection methodology compared to the collection of movement

data from social media (e.g. trajectories from Twitter, FourSquare), is that a relatively large

percentage of elderly participants (60+) were collected in our volunteer sample. This is in

contrast to assertion that GPS surveys are more likely to attract younger participants who are

probably more technologically knowledgeable than the elderly (Bricka et al. 2012). We

speculate that this might be due to the sense of inclusion: some elderly participants, who

perhaps are not comfortable with the current technological devices, may have felt that since

we made sure that the tracking task was a simple as possible, this was their one opportunity to

contribute to a technological experiment.

One of the recurring issues with the use of GPS traces for human mobility studies is the

noisiness of the trajectory data caused by the unpredictability of use of the trackers. In the

study it was assumed that participants would carry their trackers with them at all times, fully

charged and ready to track locations continuously. In practice, trackers occasionally lost

charge and were not re-charged and re-started until much later. Trackers were also taken to

unexpected locations such as destinations outside Scotland or used in unexpected travel

modes (there is a trajectory of a glider plane in our data, classified as vehicle travel by the

automatic algorithm!).

A further limitation are potential biases introduced by the short duration of the GPS survey

(one week). As we only tracked the participants over seven consecutive days, there is a

potential that we may not have captured the entire complexity of their daily movement and

socialising patterns. For example, participants could have been on an irregular schedule, on

25/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

vacation or could have experienced unusual events (emergencies or similar). Such

irregularities could introduce bias into our analysis and for individuals the frequency of visits

to places could therefore not necessarily correspond to their significance. Our solution to this

problem was to report aggregated results – the averages calculated for each group of

participants should have reduced this bias.

Another type of bias may also have been introduced with a relatively small sample size (205

individuals producing almost 4 million geocoded locations), which we tried to counter by a

relatively large number of invitations (6000) assuming a response rate of 12%. Having only

achieved a third of this response rate, this produced a smaller sample than anticipated. This

study should therefore be considered as a demonstrator for similar larger future studies rather

than providing absolute answers about movement behaviour of people in Fife.

A further issue that should be considered is the scalability of our framework to other cases

and in particular larger trajectory data sets. As trajectory analysis enters the Big Data era, it is

important that new frameworks and methods scale to increasingly larger data. Our data set is

a relatively small one and does not fit many of the characteristics of so-called Big Data (e.g.

volume, velocity, veracity, etc. (Kitchin 2013)), apart from being very fine-grained in its

resolution. Parts of the framework proposed here are based on knowing this data set very well

and it might be argued that the framework is somewhat bespoke. However, given that many

decisions performed during the development process are based on the general properties of

human movement (e.g. maximum possible velocities for driving a car, ranges accessible in

the sampling time period while moving at these velocities, the choice of movement

parameters used in neural network classification, etc.), this framework has a potential to scale

to a larger set of trajectories.

26/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Finally, there is an enormous potential of using contextually enriched GPS trajectories to

investigate a range of important social phenomena. For example, if significant places are used

as a proxy for individuals’ activity spaces as we tentatively suggest, this knowledge could be

used to improve lives of residents by identifying locations where provisions of various types

of services are inadequate causing people to travel further to fulfil their social needs. SPs and

contextually enriched trajectories could also be used to delineate and define neighbourhoods,

the boundaries of which are often contested by residents. Another possibility would be to

investigate the temporal dynamics of spatial segregation. Spatial segregation is often linked

to inequality (Palmer et al. 2013); however this phenomenon is frequently only investigated

through residential census data which provides only a snapshot. Spatial and temporal

distribution of the use of SPs linked to information on class, race or ethnicity of participants

could provide a much more complex and dynamic picture of segregation, assisting not only

social scientists, but also policy makers and urban planners addressing inequality.

In conclusion, our study is only one of examples of analyses of real human movement data

that can and will become possible in the near future. These data are now becoming readily

available through both targeted data collection efforts such as ours or through pervasive

mobile devices. Further, as we demonstrate, GPS trajectories enriched with contextual spatial

information provide the potential to better understand and model human movement

behaviour. Our study is different from previous work (reviewed in section 2) in that it

combines approaches to movement analysis from three very different areas of research. We

integrate a technological methodology (GPS tracking) with spatial analysis approaches from

computer science and transportation geography as well as with theory from sociology. Such a

combined approach provides a unique interdisciplinary perspective that would not have been

27/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

possible through the lens of a single discipline, thus opening new possibilities for the

empirical investigation of a range of social phenomena related to movement.

Acknowledgements

This work was supported by the EU FP7 Marie Curie ITN GEOCROWD grant (FP7-

PEOPLE-2010-ITN-264994). The authors would also like to thank all the survey participants.

References

Agamennoni, G, Nieto J, Nebot E, 2009 Mining GPS data for extracting significant places,

Robotics and Automation, 2009. ICRA '09. IEEE International Conference: 855-862

Anderson T, Abeywardana V, Wolf J and Lee M, 2009, National Travel Survey GPS

Feasibility Study, Department of Transport, UK.

Andrienko G, Andrienko N, Hurter C, Rinzivillo S and Wrobel S 2013 Scalable Analysis of

Movement Data for Extracting and Exploring Significant Places. IEEE Transactions in

Visualization and Computer Graphics 19(7):1078-1094

Ashbrook D and Starner T 2002 Learning Significant Locations and Predicting User

Movement With GPS, Proc. of 6th IEEE International Symposium on Wearable

Computing, Seattle, WA, 2002, 101-108

Ashbrook D and Starner T 2003 Using GPS to Learn Significant Locations and Predict

Movement across Multiple Users, Personal and Ubiquitous Computing 7: 275-286

Barron C, Neis P and Zipf A 2014 A Comprehensive Framework for Intrinsic

OpenStreetMap Quality Analysis. Transactions in GIS 18(6):877–895

28/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Beijing Municipal Committee of Transportation (BMCT) 2012 The fourth Beijing

comprehensive transportation survey report. Beijing Transport Research Centre,

Beijing Municipal Committee of Transportation.

Bhattacharya T, Kulik L and Bailey J 2012 Extracting significant places from mobile user

GPS trajectories: a bearing change based approach. Proceedings of the 20th

International Conference on Advances in Geographic Information Systems

(SIGSPATIAL '12). ACM, New York, NY, USA, 398-401

Bohte W and Maat K 2009 Deriving and validating trip purposes and travel modes for multi-

day GPS-based travel surveys: A large-scale application in the Netherlands.

Transportation Research Part C: Emerging Technologies 17(3): 285-297

Bonnel P, Lee-Gosselin M, Zmud J and Madre J-L 2009 Transport survey methods: Keeping

up with a changing world. Emerald Group Publishing, Bingley, UK.

Bricka SG, Sen S, Paleti R and Bhat CR 2012 An Analysis of the Factors Influencing

Differences in Survey-Reported and GPS-Recorded Trips. Transportation Research

Part C, 21(1):67-88

Buchin M, Driemel A van Kreveld M and Sacristan V 2011 Segmenting trajectories: A

framework and algorithms using spatiotemporal criteria. Journal of Spatial Information

Science 3: 33–63

Dodge S, Weibel R and Forootan E 2009 Revealing the physics of movement: Comparing the

similarity of movement characteristics of different types of moving objects. Computers,

Environment and Urban Systems 33: 419–434

Dodge S, Laube P and Weibel R 2012 Movement similarity assessment using symbolic

representation of trajectories. International Journal of Geographical Information

Science 26(9): 1563-1588

29/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Dunstan S, 2012, General Lifestyle Survey, Technical Appendices, Office for National

Statistics, UK.

Ewing R and Cervero R 2010 Travel and the Built Environment – A Meta Analysis. Journal

of the American Planning Association, 76(3): 265-294

Gieryn TF 2000 A Space for Place in Sociology. Annual Review of Sociology 26: 463-496

Golledge RG and Stimson RJ 1997 Spatial Behavior. The Guilford Press, New York.

Haykin SO 2008 Neural Networks and Learning Machines. 3rd edition Prentice Hall. New

York.

Hu W, Xie D and Tan T 2004 A hierarchical self-organizing approach for learning the

patterns of motion trajectories. IEEE Transactions on Neural Networks 15(1): 135–144

Huang, W, Li M, Hu W, Song G, Xing X and Xie K 2013 Cost sensitive GPS-based activity

recognition. Proceedings of the 10th International Conference on Fuzzy Systems and

Knowledge Discovery (FSKD) 2013, 962-966

Holm ED 2013 Design for solitude. In: Tjora A and Scambler G, Café Society. Palgrave

Macmillan. Chapter 9, 173-184

Isaacman S, Becker R, Caceres R, Kobourov S, Martonosi M, Rowland J and Varshavsky A

2011 Identifying Important Places in People’s Lives from Cellular Network Data.

Pervasive Computing, Lecture Notes in Computer Science, 6696:133-151

Itsubo S and Hato E 2006 A study of the effectiveness of a household travel survey using

GPS-equipped cell phones and a WEB diary through a comparative study with a paper

based travel survey. 85th Annual meeting of the Transportation Research Board,

Washington DC, USA.

Kang, JH, Welbourne W, Stewart B and Borriello G 2004 Extracting places from traces of

locations. WMASH 2004: 110–118

30/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Karlaftis, MG and Vlahogianni EI 2011 Statistical methods versus neural networks in

transportation research: differences, similarities and some insights. Transportation

Research Part C, 19(3): 387-399

Kitchin R 2013 Big data and human geography: Opportunities, challenges and risks.

Dialogues in Human Geography 3(3): 262–267

Kotz D, Henderson T and Abyzov I 2004, CRAWDAD - A Community Resource for

Archiving Wireless Data At Dartmouth. Dartmouth University, http://crawdad.org/ (last

accessed 13 Jan 2015)

Krygsman S and Nel J 2009 The use of global positioning devices in travel surveys-a

developing country application. Proceedings of the 28th annual Southern African

transport conference 2009, Pretoria, South Africa

Kwan M-P 2013 Beyond Space (As We Knew It): Toward Temporally Integrated

Geographies of Segregation, Health, and Accessibility. Annals of the Association of

American Geographers 103(5): 1078-1086

Kwan M-P and Neutens 2014 Space-time research in GIScience. International Journal of

Geographical Information Science 28(5): 851-854

Laing A and Royle J 2013 Examining chain bookshops in the context of “third place”,

International Journal of Retail & Distribution Management, 41(1):27-44

Laube P, Dennis T, Forer P and Walker Mike 2007 Movement beyond the snapshot–dynamic

analysis of geospatial lifelines. Computers, Environment and Urban Systems, 31(5):

481-501

Lian D, Xie X 2011 Learning location naming from user check-in histories Proceedings of

the 19th ACM SIGSPATIAL International Conference on Advances in Geographic

Information Systems, 112–121

31/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Liao L, Fox D and Kautz H 2007 Extracting places and activities from gps traces using

hierarchical conditional random fields. The International Journal of Robotics Research

26(1):119-134

Lin E-Y 2012 Starbucks as the Third Place: Glimpses into Taiwan’s Consumer Culture and

Lifestyles. Journal of International Consumer Marketing, 24:119–128

Long JA and Nelson T 2013 A review of quantitative methods for movement data.

International Journal of Geographical Information Science, 27(2): 292-318

Marchal P, Roux S, Yuan S, Hubert J-P, Armoogum J, Madre J-L and Lee-Gosselin M 2008

A study of non-response in the GPS sub-sample of the French national travel survey

2007-08. In P Bonnel and J-L Madre (Eds.), The 8th international conference on survey

methods in transport, Annecy, France.

Oldenburg R 1989 The Great Good Place: Cafes, Coffee Shops, Community Centers, Beauty

Parlors, General Stores, Bars, Hangouts, and How They Get You Through the Day.

New York: Paragon House

Palmer JRB, Espenshade TJ, Bartumeus F, Chung CY, Ozgencil NE and Li K, 2013, New

Approaches to Human Mobility: Using Mobile Phones for Demographic Research.

Demography 50:1105–1128

Parent C, Spaccapietra S, Renso C, Andrienko G, Andrienko N, Bogorny V, Damiani ML,

Gkoulalas-Divanis A, Macedo J, Pelekis N, Theodoridis Y and Yan Z, 2013, Semantic

Trajectories Modeling and Analysis. ACM Computing Surveys, 45(4), Article No. 42

Purves R, Laube P, Buchin M, and Speckmann B, 2014. Moving beyond the point: An

agenda for research in movement analysis with real data. Computers, Environment and

Urban Systems, 47: 1-4

32/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Rasmussen, T., Ingvardson, J. B., Halldo´rsdo´ ttir, K., & Nielsen, O. A. 2013 Using

wearable GPS devices in travel surveys: A case study in the Greater Copenhagen area.

Proceedings of the Annual Transport Conference at Aalborg University, Denmark.

Rodrigues A, Damásio C and Cunha J E 2014 Using GPS Logs to Identify Agronomical

Activities, Connecting a Digital Europe Through Location and Place. Lecture Notes in

Geoinformation and Cartography 2014: 105-121

Rodriguez DA and Joo J 2004 The relationship between non-motorized mode choice and the

local physical environment. Transportation Research Part D, 9: 151–173

Schwanen T and Moktharian PL 2005 What affects commute mode choice: neighborhood

physical structure or preferences toward neighborhoods? Journal of Transport

Geography, 13: 83–99

Sester M, Feuerhake U, Kuntzsch C and Zhang L 2012 Revealing Underlying Structure and

Behaviour from Movement Data. KI - Künstliche Intelligenz 2012: 1-9

Shaw B, Shea J, Sinha S and Hogue A 2013 Learning to rank for spatiotemporal search.

Proceedings of the 6th ACM International Conference on Web Search and Data Mining,

WSDM’13, ACM, 717–726

Steinkuehler CA and Williams D 2006 Where Everybody Knows Your (Screen) Name:

Online Games as “Third Places”. Journal of Computer-Mediated Communication 11:

885–909

Torrens P, Li X and Griffin W 2011 Building agent-based walking models by machine-

learning on diverse databases of space-time trajectory samples. Transactions in GIS

15(1): 67–94

Umair M, Kim WS, Choi BC and Jung SY 2014Discovering personal places from location

traces. Proceedings of the 16th International Conference on Advanced Communication

Technology (ICACT 2014), 709-713, 16-19 Feb. 2014

33/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Van Vugt M, Van Lange PAM and Meertens RM, 1996, Commuting by car or public

transportation? A social dilemma analysis of travel mode judgements. European Journal

of Social Psychology, 26: 373-395

Wener RE and Evans GW 2007 A Morning Stroll Levels of Physical Activity in Car and Mass

Transit Commuting. Environment and Behavior, 39(1): 62-74

Xiao X, Zheng Y, Luo Q and Xie X, 2014, Inferring social ties between users with human

location history. Journal of Ambient Intelligence and Humanized Computing, 5:3-19

Yan Z, Charkraborty D, Parent C, Spaccapietra S and Aberer K, 2013, Semantic Trajectories:

Mobility Data Computation and Annotation. ACM Transactions on Intelligent Systems

and Technology, 4(3), Article No. 49

Ye Y, Zheng Y, Chen Y 2009 Mining individual life pattern based on location history. In:

Mobile data management: systems, services and middleware, IEEE MDM’09, Taipei,

Taiwan, 18–20 May 2009, pp 1–10

Yin P, Ye M, Lee W-C, and Li Z 2014 Mining GPS Data for Trajectory Recommendation. In

Advances in Knowledge Discovery and Data Mining, Springer Verlag, Berlin-

Heidelberg, 50-61

Zheng Y, Zhang L, Xie X, and Ma W-Y 2009 Mining interesting locations and travel

sequences from GPS trajectories. Proceedings of the 18th international conference on

World wide web, WWW 2009, 791–800

Zheng Y, Chen Y, Li Q, Xie X and Ma W-Y 2010 Understanding transportation modes based

on GPS data for web applications. ACM Transactions on the Web 4(1): Article No. 1

Zhou C, Frankowski D, Ludford PJ, Shekhar S and Terveen LG 2007 Discovering personally

meaningful places: An interactive clustering approach. ACM Transactions on

Information Systems 25(3), Article No. 12

34/54
Accepted to International Journal of Geographic Information Science (Sept 2015)

35/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Tables

Table 1. Participants in the GPS survey.

Kirkcaldy Dunfermline Glenrothes

Male Age available 23 41 22
No age information 15 18 8
Total 38 59 30
Female Age available 10 23 17
No age information 7 9 12
Total 17 32 29

36/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Table 2: Grouping of original OS activity types into categories used in our analysis of places.

OS number OS activity type Category for place analysis

01 Accommodation, eating and Leisure
drinking
02 Commercial services Shopping
03 Attractions Leisure
04 Sport and entertainment Leisure
05 Education and health School or Health Care
06 Public infrastructure Leisure
07 Manufacturing and production Excluded from analysis
09 Retail Shopping
10 Transport (parking, petrol stations) Traffic stops

37/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Table 3: Average time per day (in hours) spent in each travel mode, broken down per gender,

location and day of the week (workday vs. weekday).

Day of Average time per day Kirkcaldy Dunfermline Glenrothes

the week (in hours) spent:
Male Weekday Driving 0.65 0.64 0.69
Walking 0.41 0.51 0.35
On public transport 0.18 0.07 0.22
Weekend Driving 0.91 0.76 0.71
Walking 0.66 0.46 0.46
On public transport 0.83 0.05 0.28
Female Weekday Driving 0.45 0.67 0.38
Walking 0.38 0.31 0.51
On public transport 0.09 0.26 0.08
Weekend Driving 0.48 1.24 0.69
Walking 0.66 0.46 0.46
On public transport 0.64 0.30 0.32

38/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Table 4: Accuracy of home identification

Number of Homes found within

participants 50m 100m 200m
Dunfermline 91 63 79 84
Glenrothes 59 39 55 56
Kirkcaldy 55 36 45 47
Total 205 138 179 187
Total Percentages 67.31% 87.31% 91.21%

39/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 1. Trajectories collected in the GPS travel survey – the whole extent across Scotland

and beyond (large map) and the focus on Fife and surrounding locations (insert).

40/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 2. Matching GPS points from stop locations to POIs. If the POI is a polygon

representing the actual physical object on the ground (e.g. a parking space), by taking a 50m

buffer around the polygon, all relevant points are labelled as occurring in this POI. If

however the POI is represented as a point of interest (e.g. a centroid of a parking space), then

a buffer of only 50m does not cover all relevant GPS points. A larger buffer, e.g. one of 250m

however, artificially overassigns the POI category to points that are in reality outside the POI

area (marked with red crosses).

41/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 3. The overview of our framework for identification of human mobility patterns from

trajectories. The framework is structured into three phases: separation of dynamic and static

behaviour, analysis of places and analysis of movement types. Blue rectangles mark data

input, yellow rectangles represent processing steps and green rectangles derived results

within each phase.

42/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 4. Spatio-Temporal Kernel Window statistics. a) The statistic is calculated by

counting the number of points within a neighbourhood (in this case 25m) of a specific point.

For this point the STKW value is 6. b) The difference between a point buffer which assigns

points from multiple visits to one single point (e.g. all red points within the 25m circle are

assigned to the centre point of the circle, shown in blue) and the STKW statistic, which only

counts points of the same visit (points in green). c) A typical temporal progression of the

STKW statistic with sudden changes in value when the transportation mode changes.

43/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 5. Analysis of places from stop segments. We first identify home, followed by the

most important significant places (which individuals re-visit most frequently and where they

spend the most time) and categorise these with activity types in a comparison with our Place-

Of-Interest dataset. The remaining less important stops are compared with the NaPTAN data

set in order to identify a pedestrian waiting at public transport stops or a driver stuck in

traffic. The remainder of the stops is further compared to Places-of-Interest. The two types of

traffic stops (public transport and traffic) are returned as input into the last phase of the

process, analysis of movement.

44/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 6. Average time per day (in hours) spent at home, broken down per age group and

residential location, a) on weekdays and b) on weekends.

45/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 7. Average time per day (in hours) spent in SP1-2-3, broken down per age group and

residential location, a) on weekdays and b) on weekends.

46/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 8. Small multiples (3x3) of KDE maps for SP1-2-3 vs. residential location for male

participants.

47/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 9. Small multiples (3x3) of KDE maps for SP1-2-3 vs. residential location for female

participants.

48/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Figure 10. Radar plots disaggregated by age of the average distance of SP1-2-3 from home

(shown as the central point of each plot).

49/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Supplementary Online Material – the pseudocode for algorithms used in our

framework

This document desribes the algorithms used for the identification of dynamic (travel modes)
and static (significant places) behaviour from movement data. The three-phase framework is
described in section 4 of the original paper (see figure 3 for an overview) and includes the
following algorithms:

Phase 1: Separating dynamic and static behaviour

1.1. Calculation of the Spatio-Temporal Kernel Window statistics
1.2. Trajectory segmentation
1.3. Travel mode classification & identification of stops

Phase 2: Analysis of movement

2.1. Classifying Vehicular Movement

Phase 3: Analysis of places

3.1. Extraction of home and significant places

Phase 1: Separating dynamic and static behaviour

1.1. Calculation of the Spatio-Temporal Kernel Window statistics

Input: (P, dist)

where P is a chronologically ordered set of all points of the user with
their x and y coordinates (i.e. an entire trajectory of one participant)
and dist is a threshold distance for STKW calculation

For each point in P:

Initialize counter at zero

Set current point as original point

While distance between current point and the next is less or equal to
dist:
Add one to counter
Move to next point

Return to the original point

While distance between current point and the previous is less or

equal to dist:
Add one to counter
Move to previous point

Return to the original point

Set counter as STKW value of the point

Output: updated P with STKW values for each point

50/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

1.2. Trajectory Segmentation

Input: (P, window, gap)

where P is a chronologically ordered set of all points of the user with
their x and y coordinates and STKW values; window is the number of points
taken into consideration for segmentation and gap is a distance threshold
for segmentation

Initialise set stopPointsSet

For each point in P:

If distance between current point and the next is larger than gap:
Add current point into stopPointSet

Else:

Initialise leftWindow as set of previous points of length equal to

window

Initialise rightWindow as set of next points of length equal to

window

Initialise leftIntensity as average of STKW values of points in

leftWindow

Initialise righttIntensity as average of STKW values of points in

rightWindow

Initialise difference as absolute value of difference between

leftIntensity and rightIntensity

Initialise minimum as minimum value between leftIntensity and

rightIntensity

Initialise prevDifference as -1

Initialise stopPoint as Null

While difference is larger than minimum +1:

If difference is larger than prevDifference:
Set stopPoint as current point
Set prevDifference as difference
Else:
Add stopPoint to stopPointSet
Move to next point

Divide P into segments by splitting the trajectory at each stopPoint

contained in stopPointSet

Output: the set of STKW-based segments

51/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

1.3. Travel mode classification & identification of stops

Input: (S, tSet)

where S is a set of all segments, where the median speed and median
distance to centre values were calculated across each segment; and tSet is
a prepared training set for neural network

Initialise neuralNetwork1 using tSet with ”stop” and ”movement” classes

Initialise neuralNetwork2 using tSet with ”walk” and ”vehicle” classes

For each segment in S:

Send median speed and median distance of the current segment to

neuralNetwork1

Get result1 from neuralNetwork1

If result1 equals to ”movement”:

Send median speed of the current segment to neuralNetwork2
Get result2 from neuralNetwork2
Set class of current segment to result2

For each segment in S:

If current and next segment have the same class:

If current and next segment are spatially and temporary
continuous:
Merge both segments into one

Output: classified set of segments

2. Phase 2: Analysis of movement

2.1. Classifying Vehicular Movement

Input: (S, busStops, trainStops, dist, trafficStopTime)

where S is a set of all classified segments of the user with their centroid
values and time difference between start and end of the segment; busStops
and trainStops are NaPTAN database points of public transportation stops;
dist is a distance condition for the analysis; trafficStopTime is a time
constraint for the length of traffic stops

For each segment in S:

If segment class is equal to “stop”:

Initialise busDistance as distance of the current segment’s

centroid from nearest bus stop in busStops
Initialise trainDistance as distance of the current segment’s
centroid from nearest train stop in trainStops

If the next or previous segment is classified as “vehicle”:

If trainDistance or busDistance is less than dist:

If busDistance is less than trainDistance:

Set segment’s class to “Bus Stop”
Else:

52/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Set segment’s class to “Train Stop”

Else:
If timeDifference of the segment is less than
trafficStopTime:
If the next and previous segments are
classified as “vehicle”:
Set segment’s class to “Traffic Stop”
For each segment in S:

If segment is classified as “Bus Stop”:

Initialise set busTravel

Add previous segments to busTravel until their class is not

“vehicle” or “Traffic Stop”

Add next segments to busTravel until their class is not

“vehicle” or “Traffic Stop”

Change class of all “vehicle” segments in busTravel to “Bus”

If segment is classified as “Train Stop”:

Initialise set trainTravel

Add previous segments to trainTravel until their class is not

“vehicle” or “Traffic Stop”

Add next segments to trainTravel until their class is not

“vehicle”
Change class of all “vehicle” segments in trainTravel to
“Train”

Output: a set of segments with a classified travel mode

3. Phase 3: Analysis of places

3.1. Extraction of home and significant places

Input: (S,bufferSize)
where S is a set of all classified segments of the user (output from 2.1.)
with their centroid values and time difference between start and end of the
segment; and bufferSize is a size of search area for significant places

Initialise an empty set of locations

For each segment in S:

If segment is classified as "stop":
Add current segment into the set of locations

Divide all segments in locations into groups, where segment centroids are
closer to the group average than buffer size

Sort created groups based on the number of segments they hold from those
with the largest number of segments to those with the smallest number of
segments

53/54
 Accepted to International Journal of Geographic Information Science (Sept 2015)

Initialise timeSortedGroups as a descending sort of all groups starting

with the highest number of members, based on the sum of their time
difference

Set Places as centroids of all segments within the first group of

timeSortedGroups

Output: location of the first Place

The first run of this algorithm produces the location of the home.
Subsequent runs of this algorithm are ran on an input where non-socialising
segments have been removed, to produce each of the remaining significant
places respectively (SP1, SP2, SP3).

54/54