Vrije Universiteit Brussel

Monitoring Urban-Freight Transport Based on GPS Trajectories of Heavy-Goods VehiclesHadavi, Sheida; Verlinde, Sara; Verbeke, Wouter; Macharis, Cathy; Guns, Tias

Published in:IEEE Transactions on Intelligent Transportation Systems

DOI:10.1109/TITS.2018.2880949

Publication date:2019

Link to publication

Citation for published version (APA):Hadavi, S., Verlinde, S., Verbeke, W., Macharis, C., & Guns, T. (2019). Monitoring Urban-Freight TransportBased on GPS Trajectories of Heavy-Goods Vehicles. IEEE Transactions on Intelligent Transportation Systems,20(10), 3747-3758. [8577021]. https://doi.org/10.1109/TITS.2018.2880949

General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portalTake down policyIf you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediatelyand investigate your claim.

Download date: 15. Mar. 2022

1

Monitoring Urban-Freight Transport based onGPS Trajectories of Heavy-Goods Vehicles

Sheida Hadavi, Sara Verlinde, Wouter Verbeke, Cathy Macharis, and Tias Guns

Abstract—For designing transport policy measures, it is crucialto base decisions on evidence-based insights regarding transportflows and behavior. This article introduces practical indicatorsfor urban transport, that can be derived from large collectionsof GPS trajectories of heavy-goods vehicles. The indicatorsframework enables cities, municipalities, and regions to gaininsights in the urban transport activities in their region. Wemotivate our indicators based on the objectives and action plansdescribed in the strategic plan for goods traffic of the Brussels-Capital Region in Belgium. We provide a case study on datacollected from on-board units of heavy-goods vehicles, whichbecame mandatory in Belgium as part of its dynamic road-pricing scheme in 2016. This work contributes to the excitingnew capabilities that are obtained through GPS trajectories dataand the possibilities this offers for smart cities at the tactical,operational, and strategic level.

I. INTRODUCTION

TRANSPORTATION of goods plays a central role in theeconomic activity of a city and facilitates urban lifestyle,

though it also contributes significantly to externalities causedby road transport. For example, although only 10% to 20%of urban road traffic consists of freight vehicles, they areresponsible for 16% to 50% of transport-related emission of airpollutants in cities (depending on the pollutant considered) [1],[2]. To mitigate the negative effects of urban freight transport,public and private actors take measures [3]–[5]. The challengeis to decide which objectives need to be attained, and whichmeasures to take.

A variety of collection methods for urban-freight data is inuse today, each with their own purpose and their advantagesand disadvantages: traffic counts, surveys, interviews, groupdiscussions, written questionnaires and diaries [6]–[8].

At the same time, increasingly often On-Board Units (OBU)with position tracking (GPS/GNSS) are present in freightvehicles. Despite their emerging potential for improved tripreporting [9], practical examples of GPS logging to collecturban-freight data are scarce. Comendador et al. [10] equipped20 vans with a GPS for a period of two months to study freighttraffic in two Spanish cities. More recently, Allen et al. [11]equipped drivers and vehicles of a parcel company operatingin London with GPS trackers to analyze last-mile operationsof parcel deliverers. Limitations of this type of study is that thesample of vehicles monitored is small and/or limited to onecompany. Thereby, handling big and noisy data systematically

S. Hadavi, S. Verlinde, W. Verbeke, C. Macharis and T. Guns are withMOBI (Mobility, Logistics and Automotive Technology) Research Centre,department of Business technology and Operations of vrije UniversiteitBrussel, Brussels, Belgium, corresponding e-mail: Sheida.Hadavi@vub.be(see http://mobi.vub.ac.be/about/who/).

has not been as important. Furthermore, in comparison withthis paper, there is no focus on automatically detecting stoplocations from the data, as they directly collect surveys onlocation and time of the stops from the drivers.

Increasingly though, new national or regional initiatives arebeing developed such as a national road tax per driven kilo-meter. When this data is also made accessible to the regionalor national administrations, this opens new opportunities forextracting information to gain insights into freight flows.

While there has been outstanding work on understandingpassenger mobility through GPS trajectory analysis [12]–[14],there is less research on analyzing freight flows and freight-specific properties such as parking and loading/unloadingbehavior, let alone within the context of policy-oriented in-dicators and measures.

In this work, we aim to derive quantitative measurementsfrom GPS trajectory data, received from on-board units ofheavy-goods vehicles, to understand urban freight transportin a region. We present these quantitative measurements asindicators. We guide ourselves by the Strategic Plan for Goodstraffic in the Brussels-Capital Region [15] to derive a broadset of such indicators, many of which can not be measuredwith traditional measuring methods.

Our contributions are four-fold:1) we identify a set of relevant, concrete and measurable

policy-oriented freight transport indicators that can be derivedfrom GPS trajectory data, based on an existing strategic planfor urban freight;

2) we propose a generic step-by-step approach to computethese indicators from large amounts of raw and noisy GPStrajectories that span multiple weeks;

3) we showcase the methodology on 28 days of data fromthe on-board units of all heavy-goods vehicles in Belgium, forthe territory of the Brussels-Capital Region (BCR);

4) we demonstrate that analyzing floating-car data of heavygoods vehicles can provide complementary insights and canchallenge knowledge that is based on traditional traffic counts,though data quality can hinder certain analysis.

The paper is organized as follows: in Section II we discussrelated work on urban freight indicators and data analysisof GPS trajectories. Section III motivates the set of policy-oriented freight transport indicators through the strategic planfor urban freight of the Brussels-Capital Region [15]. Then,in Section IV, we describe the step-by-step approach. InSection V we showcase our methodology on GPS trajectorydata of HGVs in BCR. Finally, Section VI highlights keydiscussion points and describes possible avenues for futurework.

2

II. RELATED WORK

A. Urban freight transport indicators

Cities have been taking various measures to enhance sus-tainability of Urban Freight Transport (UFT) on their territory[16], [17]. Quak [4] distinguishes between three types ofpolicy measures that are aimed to change UFT or at leastaffect it: road pricing, licensing and regulation, and parkingand unloading. The licensing and regulation category coversvehicle regulations, vehicle load factors, low emission zones,time restrictions and dedicated infrastructure [4]. Consideringmore recent knowledge on logistics sprawl, a fourth category,spatial planning regulations, should be added to that [18].Apart from these restrictive policy measures, cities can alsobe supportive to initiatives of private actors towards moresustainable UFT [19]. A third option is that cities initiate pilotactions to find out whether a certain solution works in practiceand achieves the envisioned goals [19], [20].

Not all cities are equally mature in terms of their freighttransport policies. Many (smaller) cities do not have a freighttransport strategy at all [21]. To improve urban freightplanning, European cities are encouraged by the EuropeanCommission to integrate UFT in their mobility plans and/orto come up with dedicated freight plans [22]. What theseplanning approaches have in common is that they require datacollection for analysis of the current mobility situation andfor monitoring the impact of measures. There still remains,however, a lack of urban freight plan evaluation [22].

Many mobility and freight plans consider monitoring asan important aspect of the planning process, but they do notprovide an adequate framework for it [23]. Since 2009, theEuropean Commission has been promoting Sustainable UrbanMobility Plans (SUMPS) [24]. Developing measurable targetsand arranging for monitoring and evaluation are distinctivesteps in SUMP processes [25]. SUMP guidelines propose toset targets and to connect them to easily-measurable indicatorsand performance measures.

Today, there is no commonly accepted indicator set for UFTthat would allow authorities to monitor how UFT changes ontheir territory over time. A methodology to evaluate urbanlogistics innovations is developed in [26]. Multiple indicatorsets for UFT exist, but they were compiled within researchprojects to for specific pilots. Many recent European H2020projects on freight transport all came up with their own set ofindicators [27]–[29]. Apart from transport indicators, all setsalso contain economic, social and environmental indicators.In these categories, conceptual indicators such as noise, pol-lution and traffic safety are introduced. How these should bemeasured in practice, and how practical this is (e.g. differentsensors across vehicles and locations) is typically left open.

For the indicator category of urban freight transport, theexisting projects have only limited indicators such as averagespeed, number of kilometers and service time. While [29]also contains the number and times of stops, they acquirethese through surveys. Moreover, analyzing locations of stops,parking and entry exit/points have not been discussed in anyof these projects [27]–[29].

B. GPS trajectory analysis

GPS data provides many opportunities for data analytics. Inthe following, we differentiate between literature investigatingthe use of GPS trajectories on passengers and trucks.

1) Human mobility: Giannotti et al. [30] defines trajectorypattern mining methods “concise descriptions of frequentbehaviors, in terms of both space and time”, and evaluatedifferent approaches. Moreover, using GPS data from vehicles,[31] uses density-based clustering algorithm to automaticallydetect significant groups of similar trips. Thereafter, somemobility characteristic of a region have been computed, witha focus on human mobility [31].

In order to get an insight in human mobility, using mobilephone data, [32] classifies users into behavioral categories of“Resident, Dynamic Resident, Commuter or Visitor”. UsingGPS data from mobile phones, [33] evaluates traffic conditionsand driving performances.

[34] identifies hazardous road locations analyzing GPSdata. Furthermore, [35] identifies activity stop locations forperson trip data using Support Vector Machines. Alternatively,[36] classifies users into returners and explorers, which bringsdiscussions around energy consumption, gas emission andurban planning. Wang et al. [37] predicts new connectionsin social networks based on mobility patterns. At a user level,[38] derives for each user a measure of mobility volume anddiversity. They aggregate the measures at municipality leveland correlate them with external socio-economic indicators.

While some indicators such as frequency of movement orlength of a trip are in common between human and freightmovements, in the case of urban-freight analysis, indicatorsto measure frequency, location and duration of parking andloading/unloading are also important due to the size and natureof these vehicles.

2) Trucks movement: A number of approaches for identify-ing a stop from raw GPS trajectories exist. Liao [39] considersa point as a stop when the computed truck speed is less than5 miles per hour and the distance traveled from the last pointis less than 100 meters. They then compare truck travel timeson different road segments in a freight corridor in the UnitedStates [39]. Similarly, [40] employs data from the Canadianowned carriers to detect stops and their purposes. They marka point as a stop when the truck has moved less than 250min 15 minutes, and its average speed has been less than 1km/h. In a second phase, they categorize a stop as primaryor secondary. If a stop is related to loading/(un)loading it isa primary stops, otherwise it has secondary purposes such asfuel refills and rest breaks [40]. Yang et al. [41] take a twostep approach to stop detection. First they use a speed of lessthan 14 km/h as a threshold to detect a stop. Thereafter, theyuse Support Vector Machine (SVM) to detect a stop from thepreliminary stops. Our stop detection is threshold-based anddiscussed in Section IV-E.

There is also increasingly research on identifying indicatorsor performance measures from GPS data at the road and citylevel. This often has a focus on travel times and congestion.Yang et al. [42] have developed a web-based informationsystem to provide travel time between selected origin and

3

destination. Moreover, they provide visualizations on the traf-fic situation based on GPS trajectories received from Scania.J. Short [43] employs data from ATRI (American TruckingIndustry) to provide travel times, and indicate the segmentsexperiencing congestion. Furthermore, they investigate effectsof closing a freeway and of a bridge collapsing. Zanjani etal. [49] also use GPS data from ATRI, to monitor truck typecomposition, and traffic flows in terms of origin-destinationmatrices.

In order to get indicators on reliability of freight flows, [44]explores GPS trajectories data to find time and location ofrecurring congestion. Similarly, [45] calculates travel timesfrom freight-specific data that are collected from GPS andautomated vehicle identification (AVI). GPS trajectories froma grocery chain have also been used to derive some perfor-mance measures such as duration, distances and number ofactivities [46], [47]. Finally, Dulebenets et al. [48] develop aninteractive ArcGIS based module for computing performancemeasures on speed, origin-destination, and information onknown parking zones and freight facilities. This requires sup-plying shapefiles with traffic zones, parking zones and otherzones of interest, while our method derives these automaticallyfrom the data for one general region of interest.

III. POLICY-ORIENTED FREIGHT TRANSPORT INDICATORS,BASED ON GPS TRAJECTORIES

We use the Strategic Plan for Urban Freight of the Brussels-Capital Region [15] to guide the identification and choice ofa relevant set of indicators that can be measured from largeamounts GPS trajectories across multiple weeks. In 2013,the government of the Brussels-Capital Region adopted itsstrategic plan for goods traffic which describes the Region’sstrategy for urban freight. This plan was co-created by theurban freight planners of the city in consultation with manydifferent stakeholders, including haulers, retailers, logisticsproviders, universities, police, trade associations and residents.Through four participative workshops, knowledge and experi-ence was shared and the plan was drafted. The strategic planand its process received the European Commission’s award forSustainable Urban Mobility Planning in 2016.

We now discuss this strategic plan in more detail, whosemain goal to have smarter and cleaner goods distribution inthe city [15]. The three main objectives are:

1) to reduce and optimize the movement of vehicles trans-porting goods in and from the region, by groupingflows whilst guaranteeing the imperatives of just-in-timedeliveries

2) to trigger a modal shift from road to waterways andrail for long distances, and encourage the use ofmore environment-friendly vehicles for last-mile jour-neys within the city

3) to improve the working conditions of deliverers, byadopting a clearer regional framework and by developingadapted infrastructures

These objectives go much further than road transport andinclude city planning, waterways and bike transports, regula-tory issues and so forth. Our focus will be on road transportby heavy-goods vehicles.

TABLE I: The 5 themes and relevant action plans (top, [15]),and indicators with between brackets the actions plans forwhich they can be used (bottom).

Theme A: Organize an urban distribution structure(1) Analyze the flows of goods in preparation for a distribution scenario

Theme B: Integrate urban distribution in land-use planning(11) Periodically complete the inventory of the logistics real estate and

compare it with the needsTheme C: Improve the efficiency of deliveries

(18) Improve road deliveries(19) Deploy itineraries for goods(20) Establish a road charge by the kilometer(21) Organize the parking of heavy goods vehicles(22) Support building sites that cause less nuisance on roads(23) Reflect on proximity delivery spaces(24) Put in place flexible delivery times(26) Limit the polluting emissions of goods traffic

Theme D: Collect data and encourage innovation(28) Monitor merchandise flows and organize counting

Theme E: Develop a favorable regional framework(33) Improve the traffic of goods exported by Brussels businesses(36) Rationalize distances of goods traffic

HGVs Movements :- number of vehicles in the region (24),(28)- distances driven in the region (20),(36)- distances per HGV (20),(36)- distances per entry (23),(36)- time of entering and leaving the region (20),(24)- entry/exit locations (1),(19),(28)HGVs Pollution:- engine types used (20),(26)HGVs Loading/Unloading:- number of stops in the region (1),(24),(28),(33)- number of stops per HGV (1),(28),(33)- number of stops per entry (1),(28),(36)- average distance among each HGV’s stops (23),(36)- origin/destination analysis (1),(28)- stop locations (1),(11),(18),(19),(22),(33)- parking locations (11),(21)

The plan contains 36 action plans, grouped around 5 themes.Table I (top) shows the 5 themes, and the action plans dealingwith road transport by heavy-good vehicles (discussed in moredetail below).

While each of the action plans has an intended effect,an open question is how this effect should be observed.Traditional counting and surveys are an option but they areexpensive and limited to specific locations, and for surveysalso specific times. However, nation-wide GPS trajectoriescollected as part of a road tax offers new possibilities.

For each of the themes and action plans in Table I (top) wenow discuss what the relevant indicators are that can be derivedfrom GPS trajectories. Table I (bottom) shows an overview ofthese indicators and the actions plans they connect to.

Theme A: Organize an urban distribution structure

The first theme in the strategic plan revolves around es-tablishing and fostering an urban distribution scenario withdistribution centers and an efficient network between them.Only the first action plan in the strategic document can bemeasured throug GPS-trajectories, namely “(1) Analyze theflows of goods in preparation for a distribution scenario” [15].The 9 other action plans revolve around distribution center de-velopment, encouraging the use of waterways and supportingpublic/private partnerships.

4

For the action plan of analyzing the flow of goods; whilewe can not observe the goods per se, we do can observe theflow of the vehicles, namely through indicators such as wherethey enter and leave the region, and how often and where theystop and what the common origins and destinations are.

Theme B: Integrate urban distribution in land-use planning

The second theme in the strategic plan is closely related tothe previous one, as managing enough good quality real-estateat strategic locations is essential for achieving an efficienturban distribution structure.

The first action in this theme is “(11) Periodically completethe inventory of the logistics real estate and compare it withthe needs”. This requires knowing where HGVs stop and howpopular the different zones are; this includes short-term stopsand long-term stops e.g. overnight parking. Such indicators,visualised through counts and maps can then be linked toknown existing infrastructures, allowing to monitor the needsand their evolution over time.

The other actions are related to city and land-use planningrather than road-based transport.

Theme C: Improve the efficiency of deliveries

This theme’s first action plan “(18) Improve road-sidedeliveries” requires indicators on where HGVs typically stop,including the lesser-popular stopping points and their relationto loading/unloading zones. Similar indicators are needed formonitoring the peculiarities of building sites “(22) Supportbuilding sites to cause less nuisance on roads”. These sitesare only temporary yet can attract a fair share of HGVs inthat period.

The action plan “(19) Deploy itineraries for goods” canbe supported by indicators on entry/exit locations, stoppinglocations and their relations. Actual road use can also bemapped back to preferred and forbidden roads for heavy-goodsvehicles, to gain insights into effectiveness and pain-points.

The action plan “(20) Establish a road charge by thekilometer” has since been established and is the reason that thedata can now be analyzed. New updates to the road charge caninclude different prices per engine class or even differentiatedpricing based on time of day. The effect of such choices canbe estimated through indicators on current use per hour andper engine class. The latter is also in support of “(26) Limitthe polluting emissions of goods traffic”.

“(21) Organize the parking of heavy goods vehicles” has afocus on longer-term parking, e.g. overnight stays. Indicatorson such locations and their popularity can be extracted fromGPS trajectories.

Finally, the action plan “(23) Reflect on proximity ofdelivery spaces” can benefit from collecting indicators on thedriven distance to and between stopping locations and thedistance per entry. The action plan “(24) Put in place flexibledelivery times” further asks for analyzing indicators on thenumber of HGVs and deliveries on a per-hour basis, to assespotential impact and shifts in delivery times.

Theme D: Collect data and encourage research and innovation

The main relevant action plan is “(28) Monitor merchandiseflows and organize counting”, including traditional count andorigin/destination analysis from existing or novel data sources.

Theme E: Develop a favorable regional framework

This theme in the strategic plan revolves around a transver-sal approach with favorable conditions to goods governance.

A role that indicators can play here is to inform themobility experts but also citizens and actors on trends infreight flows. Two relevant action plans are “(33) Improvethe traffic of goods exported by Brussels businesses” whichinvolves monitoring stopping locations and frequencies; and“(36) Rationalize distances of goods traffic” for which indica-tors on traveled distances and distances between stops can bemonitored.

Other action plans in this theme are more governanceoriented.

In this section we have discussed the 5 themes in thestrategic plan, the concrete action plans per theme that relateto heavy-good road transport, and what relevant indicators areby which the effect of these action points can be measuredthrough GPS trajectory data. Table I (bottom) provides andoverview of these indicators and the different action plans (inbrackets) for which they can be used. We next discuss how toextract these indicators from raw and noisy data GPS data.

IV. METHODOLOGY AND DATA

To analyze the data, we follow a methodology that isinspired by the CRISP-DM (Cross-industry standard processfor data mining) process [50]. Figure 1 shows the differentsteps in the methodology; we discuss each step in more detailsin the following.

A. Freight movement understanding

The aim is to better understand and monitor how heavy-goods vehicles move in a city or region, by analyzing theirGPS trajectories. The current knowledge of urban-freighttransport is based on data from traditional counting loops, sur-veys and discussions with freight service providers, inhabitantsand transport experts of the city.

B. Data Understanding

In our case, the data is collected as part of a nationalroad tax per driven kilometer for HGVs, hence the use ofOnBoard Units (OBUs) is mandatory. The main provider ofOBU devices is contractually required to share the raw data,anonymously, with the regional governments. Other companiescan also become an OBU provider, as directed by the Europeandirective on the interoperability of electronic road toll systems.There is currently no requirement for other providers to sharethe data. At the time of data collection (October 2016), themarket share of the main provider was estimated to be around95% which can be considered to be representative, althoughnot complete.

5

Fig. 1: Methodology for analyzing GPS-based freight flows

Each OBU is registered to a vehicle and its license plate,the European emission norm of the engine (Euro 0 to EuroVI) and the maximum allowed total mass including that ofany trailer it is allowed to carry.

The license plate information is not part of the anonymizeddata. Instead, the data contains the country code of the licenseplate, and a pseudo-identifier that changes every day at 2:00UTC. This is to prevent individual tracking of vehicles overlonger time spans, which is also forbidden in the license ofthe data.

The time stamp, GPS coordinates, current driving speedand direction as measured by the GPS are captured every30 seconds. The choice of 30 seconds is the data providers’decision. The device must be manually turned on by the driverand it terminates when it looses power (e.g. the engine isturned off). The data is noisy for technical reasons (malfunc-tioning devices, transmission errors, software bugs, upgrades)and human reasons (forgetting to turn on, not waiting until itis fully booted).

C. Data preparation

As input, we have a file with GPS coordinates of HGVs,and a polygon specifying the region of interest.

a) Region polygon: The movement of HGVs will beanalyzed for a specific region. This region is specified asa polygon, e.g. through a shape file. While such a polygontypically denotes the official boundaries of the region, oneshould be careful in its literal use for data analysis. Twothings can heavily distort the results of the analysis: a) thepolygon may overlap with short, disconnected pieces of roadthat are otherwise completely outside of the region. This puretransit traffic will appear as many short ’visits’ to the region.b) highways may cross the region, with much of that trafficstaying exclusively on the highway. Such highways roads arebest excluded as well, as they can dominate the counts; c)regional borders often run through the middle of a road.Because of different driving directions and noise in the GPSsignal, this will lead to vehicles appearing to frequently go inand out on that road, and to fragment the location(s) wherevehicles enter and exit.

It is hence highly recommended to review the polygon whenoverlaid on the road network, and to exclude transit roads andhighways as well as fully including partially included roads,preferably with some additional buffer.

b) GPS data: The data is made available as a CSVfile with the following columns: Pseudo-ID, X/Y coordinates,speed, direction, max. mass and Euro norm. In preprocessing,we take a database view approach where we extend the tablewith additional derived columns.

The first step is to add a column indicating if a data pointis in the region of interest or not. All data with pseudo-IDsthat have no observation in the region of interest are removed.

Next, using the time stamp and coordinates of the vehicle’sprevious observation (if any), we add columns with the dis-tance and time driven between the two data points, as well asthe speed. We also add two columns with a flag indicating ifthe vehicle entered and/or exited the region in that observation,based on the previous/next observation.

D. Modeling through aggregation

We summarize data per HGV, and separately collect allfirst/last observation of the day, entry/exit observations andstopping observations.

a) Per HGV information: We compute aggregate infor-mation per pseudo-ID, that is, over the 24 hour period in whichthe pseudo-IDs are active (in our case, 2:00 UTC to 2:00 UTCthe next day). When we provide statistics per day, we meanper such 24-hour period even though it may technically includepart of the next day in the local timezone. We also computeaggregates per hour of each 24 hour period.

Based on the indicators we wish to extract, the followinginformation is computed and stored per pseudo-ID in the giventime frame (24 hours or 1 hour), typically by aggregating overthe relevant column(s):

• the total time and distance driven in that time frame,namely the sum of the augmented time/distances betweensubsequent observations;

• the total time and distance driven in that time frame andin the region of interest;

• the area of the convex hull of all points of the HGV;• the number of times the HGV entered/exited the region;• whether the HGV started or stopped in the region of

interest;• the observed time/distances, as above but only counting

observations that are not considered a ’stop’.

b) First/last observations: This contains for eachpseudo-IDs its first and last observed coordinate and timestamp per 24-hour batch.

c) Entry/exit observations in the region: This containsall the coordinates and timestamps with pseudo-ID of entriesand exits into the region of interest.

d) Intermediary stop observations in the region: Thiscontains all the coordinates and timestamps with pseudo-ID ofestimated stops in the region. Unfortunately, estimating whenand where a vehicle has made an intermediate stop is not aneasy task.

6

0

5000

10000

15000

0 1 2 3 4 5 6 7 8 9 10Driven Distance (km)

Cou

nt

Fig. 2: Overall driven distance up to 10 km

05000

10000

0 20 40 60 80 100

120

140

160

180

200

220

240

260

280

300

Driven Distance (km)

Cou

nt

Fig. 3: Overall driven distance up to 300 km, after removingvehicles that barely move

Conceptually, if a vehicle has not transmitted data for acertain amount of time, we can faithfully assume that it hasstopped, e.g. for loading/unloading or a break. Unfortunately,there is noise in the data meaning that a vehicle sometimesdoes not transmit data even though it is driving (hence it isnot a stop), or it stopped transmitting but by the time it startstransmitting again it is hundreds of meters or more away fromits last seen location.

A first consideration is, even for stops where a vehicle didnot move, how long an HGV should have stopped/been offlinebefore considering it a potential stop. Given the size of thevehicles and the effort it takes to load and unload items fromit, we consider that all gaps in transmission of less then 5minutes are not proper stops, and they may be caused by otherreasons such as traffic light, congestion, etc. All points wherea gap of more than 5 minutes has been present are recordedto be analyzed further in post-processing.

E. Post Processing

Some peculiarities of the data require extra care to avoidmisrepresentation in the results. We take extra care withvehicles that barely move and with vehicles that drive aroundthe time of the pseudo-ID switch.

a) Vehicles that barely move: A fraction of the vehiclesdrive only a negligible distance per day. Figure 2 shows thedistribution of total distance below 10 kilometer in our dataset.It can be seen that close to 15000 HGVs drive less than 1kilometer, a considerable amount.

After investigating the trajectories of several such HGVs.In the majority of cases, the vehicle was just moved fromone parking spot to another or solely on the terrain of awarehousing center. Based on this observation, we removeHGVs that drove less than 200 meters, or less than 2 minutes,or when the convex hull area of all their points is less than500 m2. Figure 3 shows the total distance distribution afterremoving the stationary vehicles. Moreover, from the data thatis left we remove the HGVs that drives less than 100 meters inthe region of interest. This is 0.6% of additional data unused.

b) Pseudo-ID switch and first/last observations: In thecollection of all first/last observations we remove the vehiclesthat started before 2:20, or finished after 1:40 UTC. We cannot be sure that this was their first or last observation of the

0 100

200

300

400

500

600

700

800

900

1000

1100

1200

1300

1400

1500

Driven Distance (m)

0 1 2 3 4 5 6 7 8 9 10Speed (km/h)

Fig. 4: Stops longer than 5 minutes: (top) distance traversed,(bottom) average speed.

day, as they may have been driving and/or made a stop around2:00 UTC, the time of the pseudo-ID switch.

c) Stops: For the candidate stopping points, Fig. 4 showthe distance driven in meters and the average velocity duringthe stop. Considering these figures we only consider a vehicleto have properly stopped if it drove less then 300 meters andwith velocity below 1.8 km/h on average. These vehicles havemoved too much without being observed, which means theirstopping location can not be sufficiently determined for furtheranalysis.

F. Visualization

Apart from standard box plots and bar charts, we use threevisualizations that aim to maximize informativeness: hourlymeans with quantiles, location analysis with heat maps andcounts, and origin/destination with relative heat matrix.

a) Hourly means with quantiles: When showing hourlystatistics, e.g. number of vehicles driving per hour as shown inFig. 7, instead of just showing mean data we also visualize thequantiles as a band around it: the median as a solid line, the25%-75% quantiles as a band and the 5%-95% quantile dataas a light band. This provides an impression of the varianceof the data without distracting with outliers.

b) Location analysis with heat maps and counts: To dolocation analysis, we opt to use high resolution heat maps,as shown in Fig. 15. A determining parameter here is thebandwidth used when computing the density estimates, whichinfluences the size of the regions. We decreased the bandwidthuntil the regions were sufficiently non-overlapping. For eachof the regions we extract the outer polygon and count thenumber of observations in it. This is used to rank the differentregions and show their relative importance.

We do not rely on the statistical regions as used for officialdemographic data, as these regions are still fairly large anddo not pinpoint the true location of activity. We also do notuse a grid approach as often done, because this can split thelocations in pieces and hence fragment the counts.

c) Origin/destination with relative heat matrix: Origindestination matrices are a traditional tool in mobility analysis.To analyze the flows of vehicles in the region in terms ifentry and exit points, we visualize the entry/exit relations asa colored matrix as shown in Fig. 14 where each column isan entry zone and each row-element shows the percentageof vehicles that exit at the corresponding zone. This aidsinterpretation of in/out versus transit behavior for each of thezones.

The analysis is done in R, and the code is available athttps://github.com/shadavi/monitoring-HGVs-GPS.git.

7

Fig. 5: Map of Belgium and its main highways

050000

100000150000

Days

Cou

nt

Others Ring BCR

Fig. 6: Number of HGVs in BCR, Ring and Others

V. ANALYSIS IN THE BRUSSELS-CAPITAL REGION

In this case study, we analyze data from 4 weeks (28days) in October 2016. The dataset contains data from themovement of all vehicles in Belgium, which amounts to 4million pseudo-IDs. The pseudo-IDs change every 24 hoursand do not represent total unique vehicles. The analysis hasbeen done with as region the Brussels-Capital Region (BCR).Figure 5 shows Belgium in light gray with BCR in dark greyand the main highways highlighted (part of the ring is notclassified highway). We modified the shapefile of the regionfollowing the description in the methodology to exclude thering road and transit roads, and to include partially includedborder roads.

Figure 6 shows the number of HGVs that drive per dayfrom October 1st to 28th 2016 in Belgium, on BCR’s part ofthe ring road but not within the BCR, and in BCR. We seethat on average around 180 000 HGVs drive in Belgium perworking day, 10 000 of which drive in BCR. On average, fromall vehicles driving in Belgium 5.5% drive within BCR for aperiod of the day, which are the vehicles we will analyze inthis research.

In the following, each of the indicators discussed in SectionIII are investigated. We do this with data from the weekdaysonly, as these are of primary interest.

A. HGVs Movements inside BCR

a) Number of Vehicles in the region; Number of kilome-ters driven in the region: In Fig. 7 and 8, we look at thenumber of vehicles driving and total number of kilometersdriven in the region during each hour of the working daysrespectively. These plots can be used to observe the counts andtotal distance driven by HGVs within the region in differenthours of the day. This can be tracked monthly or yearly to seechanging trends.

The counts are per hour, so any time between two hours isrounded to the hour before, e.g. anytime between 6:00:00 to6:59:59 is rounded to 6. The numbers are the daily average,with the percentile bands show the variations over differentweekdays. They are not clearly visible because there is barelyany variation in the total counts during these 4 weeks. We can

0

1000

2000

3000

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Cou

nt

25 -- 75%Quantile

5 -- 95%Quantile

Median

Fig. 7: Number of HGVs driving in BCR

0

5000

10000

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Tota

l driv

en

dist

ance

(km

)

25 -- 75%Quantile

5 -- 95%Quantile

Median

Fig. 8: Number of total KMs driven in BCR

0.0%10.0%20.0%

0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100

% o

f HG

Vs

0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100

Driven distance in BCR (km)

Fig. 9: HGVs driven distance in BCR per day

0.0%10.0%20.0%30.0%

0 5 10 15 20 25 30 35 40 45 50 55 60 65 70

% o

f HG

Vs

0 5 10 15 20 25 30 35 40 45 50 55 60 65 70Driven distance in BCR per entry (km)

Fig. 10: Driven distance in BCR per entry

observe that from 6:00 to 15:00, from 2 000 to 3 000 HGVsare driving per hour in BCR, and more than 9 000 KM perhour is driven. Furthermore, from hours 5 to 7, we observethe biggest increase in number of HGVs.

b) Number of kilometers per HGV; Number of kilometersper entry: Fig. 9 and Fig. 10 are histograms of driven distancesin BCR by each HGV and per entry in BCR respectively.The numbers show that 75% of the HGVs drive below 20kilometers (less than 10 kilometers on average) per 24 hoursin BCR as a whole, and 10 kilometers (less than 5 kilometerson average) per entry.

c) Time of entering and leaving the region; Entry/exitzones : Figure 11 and 12 provide an overview on the numberof HGVs entering and leaving BCR across all borders perhour. We observe a decrease in number of entries in themorning during rush hours. From 10:00 on, the number ofHGVs leaving BCR (Fig. 12) exceeds the number of HGVsentering BCR. It explains the observed decrease in number ofHGVs driving around within the Region (Fig. 7).

In Fig. 13, we identify the popular entry and exit zones

8

0

500

1000

1500

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Cou

nt

25 -- 75%Quantile

5 -- 95%Quantile

Median

Fig. 11: Number of HGVs entering BCR

0

500

1000

1500

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Cou

nt

25 -- 75%Quantile

5 -- 95%Quantile

Median

Fig. 12: Number of HGVs leaving BCR

Fig. 13: Entry and exit zones(Zone14: 17.3%, and Zone4: 14.7% are the most popular)

Entry Zone

Exit

Zone

123456789

1011121314

1 2 3 4 5 6 7 8 9 10 11 12 13 14

0.0

0.2

0.4

0.6

0.8

1.0

Fig. 14: Percentage at entry zone that leave by exit zone

of the region. With floating car data not only can we countthe entries and exits, but we can also group the entry andexit points of the vehicles. To complement this, Figure 14shows how often each exit zone (Y-axis) is used by the HGVsentering the BCR from each entry zones (X-axis). We clearlyobserve that most HGVs enter and leave the region throughthe same entry/exit point, with more interaction between zones8-10.

d) Parking zones: Figure 15 depicts the automaticallydetected zones in BCR, used by HGVs for longer term parking.These have been extracted by taking for each HGV the first andlast point observed after/before a buffer around the pseudo-ID switch (see Sec. IV-E). Then, the densities are computedand visualized through a heatmap. The automatically detectedzones correspond to known parkings, warehouses and marketsthat allow overnight staying. For sensitivity reasons, we omit a

Fig. 15: Regions HGVs start/stop their drive (parking)

Belgian

Others

0.00 0.25 0.50 0.75 1.00Count

N/AEuro6

Euro5Euro4

Euro3Euro2

Euro1Euro0

Fig. 16: Overview on Euroclass categories of HGVs

0500

1000150020002500

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Cou

nt

25 -- 75%Quantile

5 -- 95%Quantile

Median

Fig. 17: Number of HGVs stopping in BCR

more detailed discussion of the numbers and relations betweenthe zones.

B. HGVs Pollution

Based on the emission standards of the engine, each vehiclein Europe has a euro level associated with it, euro 6 beingthe best and euro 0 the worst. Figure 16 shows a comparisonbetween the average number of Belgian vs non-Belgian HGVsthat drive in BCR, and the Euro level associated with them. Wesee that most of the analyzed HGVs are Belgian. Moreover,a higher percentage of non-Belgian HGVs have euro level 5and 6 in comparison to Belgian vehicles, which generally haveless clean engines.

More fine-grained analysis of emissions requires knowingwhether the HGV had a trailer or not and what its load was.Such information is not available in the data and we hencecan not reliably perform such analysis.

C. HGVs Loading/Unloading

a) Number of Stops in the region; Number of stops perHGV ; Number of stops per entry: Figure 17 shows thenumber of stops per hour of the day. The definition of astop, as explained in the methodology, is when a HGV isnot submitting data for a period of longer than 5 minutes,during which the HGV has moved less than 300 meters witha speed below 1.8km/h. We observe a high number of vehiclesstopping in BCR between 6 and 16, with its peak being at 8 to13 (Fig. 17), following the general trend of number of vehiclesin Brussels (Fig. 7).

9

0.0%10.0%20.0%30.0%

0 1 2 3 4 5 6 7 8 9 101112131415%

of H

GVs

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Number of stops per HGV

Fig. 18: Number of loading/unloading stops in BCR

0.0%10.0%20.0%30.0%40.0%50.0%

0 1 2 3 4 5 6 7 8 9 10 11

% o

f HG

Vs

0 1 2 3 4 5 6 7 8 9 10 11Number of stops per entry

Fig. 19: Number of loading/unloading stops per entry

Origin

Des

tinat

ion

1

2

3

4

5

6

7

8

9

10

11

1 2 3 4 5 6 7 8 9 10 11

0.810.020.010.01

000

0.020.080.010.01

0.050.760.010.01

000

0.020.05

00.01

0.040.020.73

00.01

00.010.020.050.020.02

0.050.030.010.710.02

00

0.020.050.010.01

0.080.020.020.020.660.010.010.020.050.010.01

0.130.050.030.010.040.340.050.050.080.030.03

0.060.020.05

00.030.030.6

0.020.050.040.02

0.040.020.01

0000

0.80.05

00.02

0.080.020.010.010.01

00

0.020.8

0.010.01

0.030.010.050.010.01

00.020.010.060.740.01

0.040.020.020.010.010.010.010.050.040.010.72

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Fig. 20: Origin vs. Destination of HGVs

Figure 18 and 19 show the number of stops per HGV andper entry respectively. In order to avoid counting queuingHGVs, if a HGV has two consecutive stops in a distance ofless than 100 meters, we count it only once. We can see that75% of HGVs have less than 2 deliveries and less than 1delivery per hour.

b) Origin and Destination: Using information on thedifferent provinces, Fig. 20 and Table II provide an overviewon origin and destination provinces for HGVs that drove inBCR. In Fig. 20, we cross-match the origin and destinationof each HGV. It shows that most vehicles start and end theirtrip in the same province, except the province of Luxembourg.Most of HGVs entering or leaving from Luxembourg, do notgo back (on the same day). More than 50% of HGVs that cometo BCR have origin and destination in province of FlemishBrabant and Brussels-Capital Region.

c) Average distance between stops: Figure 21 shows theaverage distance between stops of the same HGV. We cansee that 75% of the HGVs drive less than 4 kilometer beforegetting to their next stopping point.

d) Loading/unloading zones: Fig. 22 shows a densitymap of the most popular zones, with the top ten zones marked.We omit concrete numbers and more details due to sensitivity.

VI. DISCUSSION

GPS trajectories of HGVs offer new capabilities for moni-toring freight flows in a region. Similar performance measures

TABLE II: Origin and Destination of HGVs

Region Orig % Dest %BCR 23.94 24.27Antwerp 9.77 9.67Hainaut 6.38 6.16Limburg 3.04 2.91Liege 2.45 2.38Luxembourg 0.58 0.59Namur 1.41 1.44East Flanders 10.18 10.07Flemish Brabant 27.56 26.73Walloon Brabant 4.40 4.24West Flanders 3.38 3.74NA 6.92 7.80

0.0%2.5%5.0%7.5%

10.0%12.5%

0 1 2 3 4 5 6 7 8 9 10 11

% o

f HG

Vs

0 1 2 3 4 5 6 7 8 9 10 11Average distance between stops (km)

Fig. 21: Average distance between stops

Fig. 22: Heat map of stopping points in BCR with top 10marked

050000

100000150000

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Tota

l driv

en

dist

ance

(km

)

25 -- 75%Quantile

5 -- 95%Quantile

Median

Fig. 23: Total kilometers on all roads in Belgium, subset ofHGVs that visit BCR at some point during the day.

can be extracted compared to fixed measurement points suchas number of entry/exits at specific locations and distancesand hourly trends on measured segments. However, manycomplementary indicators can be extracted as well: distancesper HGV over the entire region, number and location of stopsand parking, automatic detection of high-activity zones, etc.We discuss a number of possibilities in the following.

a) Driven kilometers vs. number of vehicles: ComparingGPS-based results to previous analysis based on traffic counts[51], we can analyze certain phenomenon in more detail.Analyzing the number of driven kilometers per hour at nationalscale leads to figures like Fig. 23, which we generated fromour data. The dip in number of kilometers during the morningrush hours (7-9) creates the impression that fewer HGVs are

10

0200040006000

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Cou

nt

25 -- 75%Quantile

5 -- 95%Quantile

Median

Fig. 24: Total number of HGVs on all roads in Belgium, subsetof HGVs that visit BCR at some point during the day.

driving at that time. However, with our data we can countthe number of unique vehicles on the network, rather thanhaving to derive this from the total number of kilometersdriven. Fig. 24 challenges the previous assumption and showsthat, at the national level, while the biggest increase in HGVsis indeed before 6-7, the total number of trucks continuesto increase until 10-11 o’clock. This is not evident whenlooking at number of kilometers because the average drivingspeed decreases due to congestion, which is mainly caused bypassenger vehicles.

b) Coordinate-based versus road-based analysis: Weused a coordinate-based approach for the indicators, includinglocation analysis, e.g. for the entry/exit points, as well asfor identifying stopping and parking hotspots. An alternativeapproach is to perform map matching and to match eachobservation to a specific road. Such road-based analysis iscommon [39], [43] and complements the analysis methodsdescribed here.

For generating heatmaps we use density-based locationanalysis. This requires some tuning of the parameters to obtainreasonably small clusters. Other techniques exist for locationanalysis such as density-based clustering [52], which wasnot able to sufficiently fine-grained clusters, and grid-basedapproaches, which cut the regions of interest at arbitrarilypoints thereby fragmenting the counts, while statistical regionsare too large.

c) Follow-up tools: Having real-time data, the indicatorsdiscussed can be used as a tool to evaluate the evolution ofheavy-good freight transport over time. In Fig. 25, Fig. 26,and Fig. 27, we analyze data from one year later, namely thelast two weeks of October 2017. Fig. 25 shows the numberof HGVs, which can be compared with Fig. 7 from 2016.Furthermore, the dashed line in Fig. 25 indicates the numberof HGVs on November 1st 2017, which is a weekday but aholiday in Belgium. Fig. 26 demonstrates the number of HGVsstopping. The dashed line shows the median of same periodin 2016. We can see that the number of stops has decreased in2017. Fig. 27 presents the most dense places used by trucksfor their parkings, which can be compared with Fig. 15 from2016. The increased importance of region nr. 1 in Fig. 27 isdue to a large distribution center that opened in that regionand which had activity in the South-West of BCR before.

d) Stop detection accuracy: Accurately detecting whenand where a vehicle has stopped is not trivial, due to noisein the data, the 30-second granularity and other factors likeslow-to-start OBUs and forgetful drivers. At the same time,the detection and analysis of stops is one of the main novelties

0

1000

2000

3000

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Cou

nt

25 -- 75%Quantile

5 -- 95%Quantile

1Nov2017 Median

Fig. 25: Number of HGVs driving in BCR (2017)

0500

100015002000

0 2 4 6 8 10 12 14 16 18 20 22Hour of the day

Cou

nt

25 -- 75%Quantile

5 -- 95%Quantile Median Oct 2016,

Median

Fig. 26: Number of HGVs stopping in BCR (2017)

Fig. 27: Regions HGVs start/stop their drive (parking)(2017)

that this type of data has to offer.The approach to stop detection we took in Section IV has

been rather conservative: vehicles that moved more than 300meters while the devices were off were not considered tobe stopping. This means that the locations obtained from thestopping points remaining will be reasonably close to the truestopping point, which is good for the location analysis. Italso means that we may disregard stops whose observationin the data is noisy, and hence that our counts are under-approximations. Good stopping detection is also a barrier todetecting and quantifying on-road stopping behaviour (doubleparking) and the use of loading/unloading zones and a topicof further research.

e) Data limits: While our data contains the euro emissionlevel of the HGV’s engine, the maximum allowed total mass(including optional trailer) is not indicative of the true vehiclesize, and we have no information on the loading factoreither. Especially without data on the latter, emissions andenvironmental impact are hard to assess. It should also benoted that HGV’s contain not just traditional freight vehiclesbut also vehicles like garbage trucks and construction vehicles.No information on the intended use of the vehicle, haulingof construction goods, containers, cars, chilled goods, ... isavailable either.

Finally, while HGVs are a very visible type of freightvehicle, there are many more medium to light weight trucksand vans that are used for transporting goods. However, theseare not captured in the data of our case study. The analysismethods and indicators are equally valid for GPS trajectories

11

of such freight vehicles though.

VII. FUTURE WORK

Based on a set of concrete action points, we have focusedon identifying a broad set of indicators that give a globalview on urban freight transport by HGVs. For each of theaction points discussed in Section III one could devise more in-depth or fine-grained types of analysis to do, such as verifyingrecommended road use or official loading/unloading zone use,which we did not address in this paper. Trajectory patternmining could be used to provide a more detailed picture onflows, e.g. directional behavior or changes in driving patterns.

A key point to improve is the stop detection, where perhapsmore of the context of the vehicle’s movement should be takeninto account: what driving pattern did it have before/after,where is it, etc. It would also be interesting to differenti-ate between types of visiting behavior: what proportion istransit traffic, or do they have a few or many stops, andwhat are further differences. Another interesting point forfuture research is activity detection to discriminate betweenconstruction vehicles, garbage trucks, retail suppliers, etc.

Finally, the current indicators provide a snapshot view ofthe current situation given a set of data. This can be used togenerate weekly or monthly reports. However, in such a settingthe differences between the current period and the previousperiod also play an important role. Such a more discriminativesetting, with a focus on automatically detecting trends andchanges is another avenue of future work.

ACKNOWLEDGMENT

This work was supported by the European Regional De-velopment Fund (ERDF) and the Brussels-Capital Region-Innoviris within the framework of the Operational Programme2014-2020 through the ERDF-2020 project ICITY-RDI.BRU

REFERENCES

[1] L. Dablanc, “Goods transport in large european cities: Difficult toorganize, difficult to modernize,” Transportation Research Part A: Policyand Practice, vol. 41, no. 3, pp. 280–285, 2007.

[2] S. Verlinde, “Promising but challenging urban freight transport solutions:freight flow consolidation and off-hour deliveries,” PhD diss., VrijeUniversiteit Brussel and Universiteit Gent, 2015.

[3] M. Browne, M. Piotrowska, A. G. Woodburn, and J. Allen, “Literaturereview wm9: Part i-urban freight transport, carried out as part of workmodule 1, green logistics project,” 2007.

[4] H. Quak, Sustainability of urban freight transport: Retail distributionand local regulations in cities, 2008, no. EPS-2008-124-LIS.

[5] C. Macharis and B. Kin, “The 4 a’s of sustainable city distribution:Innovative solutions and challenges ahead,” International Journal ofSustainable Transportation, vol. 11, no. 2, pp. 59–71, 2017.

[6] M. Browne, J. Allen, M. Browne, and J. Allen, “D 3.1 best practice indata collection, modelling approaches and application fields for urbantransport models i,” Theme: Urban freight data collection–synthesisreport, Best urban Freight Solutions II–EU-FP6 Co-ordination Action,2006.

[7] J. Allen and M. Browne, “Review of survey techniques used in urbanfreight studies,” Project Report of the Green Logistics: Work Module,vol. 9, 2008.

[8] C. Ambrosini, D. Patier, and J.-L. Routhier, “Urban freight establishmentand tour based surveys for policy oriented modelling,” Procedia-Socialand Behavioral Sciences, vol. 2, no. 3, pp. 6013–6026, 2010.

[9] E. Murakami and D. P. Wagner, “Can using global positioning system(gps) improve trip reporting?” Transportation research part c: emergingtechnologies, vol. 7, no. 2, pp. 149–165, 1999.

[10] J. Comendador, M. E. Lopez-Lambas, and A. Monzon, “A gps analysisfor urban freight distribution,” Procedia-Social and Behavioral Sciences,vol. 39, pp. 521–533, 2012.

[11] J. Allen, M. Piecyk, M. Piotrowska, F. McLeod, T. Cherrett, K. Ghali,T. Nguyen, T. Bektas, O. Bates, A. Friday et al., “Understanding theimpact of e-commerce on last-mile light goods vehicle activity in urbanareas: the case of london,” Transportation Research Part D: Transportand Environment, 2017.

[12] M. Lin and W.-J. Hsu, “Mining gps data for mobility patterns: A survey,”Pervasive and Mobile Computing, vol. 12, pp. 1–16, 2014.

[13] L. Shen and P. R. Stopher, “Review of gps travel survey and gps data-processing methods,” Transport Reviews, vol. 34, no. 3, pp. 316–334,2014.

[14] S. Dabiri and K. Heaslip, “Transport-domain applications of widelyused data sources in the smart transportation: A survey,” arXiv preprintarXiv:1803.10902, 2018.

[15] B. Mobility, “Strategic plan for goods transport in Brussels-CapitalRegion,” Brussels Mobility, Tech. Rep., 2013. [Online]. Available:https://mobilite-mobiliteit.brussels/en/node/194

[16] R. Ducret, D. Diziain, and T. Plantier, “Proposal for an evaluation gridfor analysing local public urban freight policies: strengths, weaknessesand opportunities for french cities,” Transportation Research Procedia,vol. 12, pp. 105–118, 2016.

[17] J.-P. Rodrigue, “Transport geography should follow the freight,” Journalof Transport Geography, vol. 14, no. 5, pp. 386–388, 2006.

[18] K. Aljohani and R. G. Thompson, “Impacts of logistics sprawl on theurban environment and logistics: Taxonomy and review of literature,”Journal of Transport Geography, vol. 57, pp. 255–263, 2016.

[19] P. Lebeau, S. Verlinde, C. Macharis, and J. Van Mierlo, “How canauthorities support urban consolidation centres? a review of the ac-companying measures,” Journal of Urbanism: International Researchon Placemaking and Urban Sustainability, pp. 1–19, 2017.

[20] M. Lindholm, “A sustainable perspective on urban freight transport:Factors affecting local authorities in the planning procedures,” Procedia-Social and Behavioral Sciences, vol. 2, no. 3, pp. 6205–6216, 2010.

[21] M. Lindholm, “How local authority decision makers address freighttransport in the urban area,” Procedia-Social and Behavioral Sciences,vol. 39, pp. 134–145, 2012.

[22] K. Fossheim and J. Andersen, “Plan for sustainable urban logistics–comparing between scandinavian and uk practices,” European TransportResearch Review, vol. 9, no. 4, p. 52, 2017.

[23] H. Buldeo Rai, T. van Lier, D. Meers, and C. Macharis, “An indicatorapproach to sustainable urban freight transport,” Journal of Urbanism:International Research on Placemaking and Urban Sustainability, pp.1–22, 2017.

[24] A. D. May, “Encouraging good practice in the development of sustain-able urban mobility plans,” Case studies on transport policy, vol. 3,no. 1, pp. 3–11, 2015.

[25] F. Wefering, S. Rupprecht, S. Buhrmann, and S. Bohler-Baedeker,“Developing and implementing a sustainable urban mobility plan,”European Platform on Sustainable Urban Mobility Plans (SUMP),Tech. Rep., January 2014. [Online]. Available: http://www.eltis.org/nl/guidelines/sump-guidelines

[26] D. Patier and M. Browne, “A methodology for the evaluation of urbanlogistics innovations,” Procedia-Social and Behavioral Sciences, vol. 2,no. 3, pp. 6229–6241, 2010.

[27] S. Verlinde and B. kin, “Citylab deliverable d5.1. definition ofnecessary indicators for evaluation,” CITYLAB – City Logistics inLiving Laboratories, Tech. Rep., November 2015. [Online]. Available:http://www.citylab-project.eu/deliverables/D5 1.pdf

[28] E. Nathanail, G. Adamos, L. Mitropoulos, M. Gogas, I. Karakikes,S. Iwan, and M. Kiba-Janiak, “Integrated assessment framework foruft solutions,” novelog, Tech. Rep., April 2016. [Online]. Available:http://novelog.eu/downloads-2/downloads/

[29] F. Ferrero, C. Guerlain, and T. Schwartz, “Kpis and methodologiesfor construction logistics,” Sustainable Urban Consolidation CentrESfor conStruction (SUCCESS), Tech. Rep., October 2015. [Online].Available: http://www.success-urbanlogistics.eu/project-knowledge/

[30] F. Giannotti, M. Nanni, F. Pinelli, and D. Pedreschi, “Trajectory patternmining,” in Proceedings of the 13th ACM SIGKDD international con-ference on Knowledge discovery and data mining. ACM, 2007, pp.330–339.

[31] F. Giannotti, L. Gabrielli, D. Pedreschi, and S. Rinzivillo, “Understand-ing human mobility with big data,” in Solving Large Scale LearningTasks. Challenges and Algorithms. Springer, 2016, pp. 208–220.

12

[32] L. Gabrielli, B. Furletti, R. Trasarti, F. Giannotti, and D. Pedreschi, “Cityusers’ classification with mobile phone data,” in Big Data (Big Data),2015 IEEE International Conference on. IEEE, 2015, pp. 1007–1012.

[33] J. Goncalves, J. S. Goncalves, R. J. Rossetti, and C. Olaverri-Monreal,“Smartphone sensor platform to study traffic conditions and assessdriving performance,” in Intelligent Transportation Systems (ITSC), 2014IEEE 17th International Conference on. IEEE, 2014, pp. 2596–2601.

[34] N. Agerholm, R. Jensen, and C. S. Andersen, “Identification of haz-ardous road locations on the basis of jerks,” in ICTCT, 2015.

[35] L. Gong, H. Sato, T. Yamamoto, T. Miwa, and T. Morikawa, “Identi-fication of activity stop locations in gps trajectories by density-basedclustering method combined with support vector machines,” Journal ofModern Transportation, vol. 23, no. 3, pp. 202–213, 2015.

[36] L. Pappalardo, F. Simini, S. Rinzivillo, D. Pedreschi, F. Giannotti, andA.-L. Barabasi, “Returners and explorers dichotomy in human mobility,”Nature communications, vol. 6, 2015.

[37] D. Wang, D. Pedreschi, C. Song, F. Giannotti, and A.-L. Barabasi,“Human mobility, social ties, and link prediction,” in Proceedings of the17th ACM SIGKDD international conference on Knowledge discoveryand data mining. ACM, 2011, pp. 1100–1108.

[38] L. Pappalardo, D. Pedreschi, Z. Smoreda, and F. Giannotti, “Using bigdata to study the link between human mobility and socio-economicdevelopment,” in Big Data (Big Data), 2015 IEEE International Con-ference on. IEEE, 2015, pp. 871–878.

[39] C.-F. Liao, “Using archived truck gps data for freight performanceanalysis on i-94/i-90 from the twin cities to chicago,” 2009.

[40] K. Gingerich, H. Maoh, and W. Anderson, “Classifying the purposeof stopped truck events: An application of entropy to gps data,”Transportation Research Part C: Emerging Technologies, vol. 64, pp.17–27, 2016.

[41] X. Yang, Z. Sun, X. Ban, and J. Holguın-Veras, “Urban freight deliverystop identification with gps data,” Transportation Research Record:Journal of the Transportation Research Board, no. 2411, pp. 55–61,2014.

[42] C. Yang, X. Ma, and Y. Ban, “Demonstration of intelligent transportapplications using freight transport gps data,” in Transportation ResearchBoard 95th Annual Meeting, 2016.

[43] J. Short, “Truck gps data for tracking freight flows,” TransportationResearch E-Circular, no. E-C187, 2014.

[44] M. E. Hallenbeck, E. McCormack, J. Nee, and D. Wright, “Freight datafrom intelligent transportation system devices,” The Center, Tech. Rep.,2003.

[45] F. Srour and D. Newton, “Freight-specific data derived from intelligenttransportation systems: Potential uses in planning freight improvementprojects,” Transportation Research Record: Journal of the Transporta-tion Research Board, no. 1957, pp. 66–74, 2006.

[46] K. Sturm, Z. Pourabdollahi, A. Mohammadian, and K. Kawamura, “Gpsand driver log-based survey of grocery trucks in chicago, illinois,”Transportation Research Record: Journal of the Transportation ResearchBoard, no. 2410, pp. 31–38, 2014.

[47] X. Yang, Z. Sun, X. J. Ban, J. Wojtowicz, and J. Holguın-Veras, “Urbanfreight performance evaluation using gps data,” in Annual Meeting,Washington, vol. 600, 2014, p. 45.

[48] M. A. Dulebenets, K. Pujats, and N. Deligiannis, “Development of toolsfor processing truck gps data and analysis of freight 1 transportationfacilities 2,” Transportation, vol. 2, p. 3, 2017.

[49] A. B. Zanjani, A. R. Pinjari, M. Kamali, A. Thakur, J. Short, V. Mysore,and S. F. Tabatabaee, “Estimation of statewide origin–destination truckflows from large streams of gps data,” Transportation Research Record:Journal of the Transportation Research Board, vol. 2494, pp. 87–96,2015.

[50] P. Chapman, J. Clinton, R. Kerber, T. Khabaza, T. Reinartz, C. Shearer,and R. Wirth, “Crisp-dm 1.0 step-by-step data mining guide,” TheCRISP-DM consortium, Tech. Rep., August 2000. [Online]. Available:http://www.crisp-dm.org/CRISPWP-0800.pdf

[51] P. Lebeau and C. Macharis, “Freight transport in brussels and its impacton road traffic?” Brussels Studies. La revue scientifique electroniquepour les recherches sur Bruxelles/Het elektronisch wetenschappelijktijdschrift voor onderzoek over Brussel/The e-journal for academicresearch on Brussels, 2014.

[52] M. Ankerst, M. M. Breunig, H.-P. Kriegel, and J. Sander, “Optics:ordering points to identify the clustering structure,” in ACM Sigmodrecord, vol. 28, no. 2. ACM, 1999, pp. 49–60.

Sheida Hadavi received a B.S. in Computer Engi-neering from K. N. Toosi University of Technologyin 2006, and a M.S. in Computer Science from VrijeUniversiteit Brussel in 2011. Since 2012 she is aPh.D. researcher in MOBI (Mobility, Logistics andAutomotive Technology) Research Centre, BUTO(Business Technology and Operations) department,in Vrije Universiteit Brussel. Her research focuseson data analytics in sustainable transport.

Sara Verlinde received a M.S. in Architecture En-gineering in 2003 at Gent University, and a Ph.D. in“Freight flow consolidation concepts and off-hourdeliveries” from Economics an Geography facultiesof Vrije Universiteit Brussel and Ghent Universityin 2015. Today, she is a postdoctoral research as-sociate in the research group MOBI, where she isresponsible for the City Logistics team. On-goingmajor projects include Civitas Citylab (H2020 –City Logistics in Living Laboratories) and a nationalresearch project on off-hour deliveries.

Wouter Verbeke is an assistant professor of busi-ness informatics and data analytics at Vrije Uni-versiteit Brussel. He graduated in 2007 as a CivilEngineer and obtained a Ph.D. in applied economicsat KU Leuven in 2012. His research is mainlysituated in the field of predictive, prescriptive, andnetwork analytics, and is driven by real-life busi-ness problems that require a data driven solution,including applications in marketing, finance, as wellas transport and mobility. Wouter teaches severalcourses on information systems and advanced mod-

eling for decision making to business students, and has tutored several courseson credit risk modeling and customer analytics to business professionals.

Cathy Macharis received a M.S. in Commer-cial Engineering from Vrije Universiteit Brussel in1994, and a Ph.D. in “Strategic modeling for intermodal terminals” from faculty of Applied EconomicSciences-Commercial Engineering of Vrije Univer-siteit Brussel in 2000. She is a full professor in VrijeUniversiteit Brussel. She is the head of the researchgroup MOBI and the BUTO department. She isspecialized in the assessment of policy measuresand innovative concepts in the field of sustainablelogistics and urban mobility.

Tias Guns Tias Guns received a M.S. in Com-puter Science in 2006 at the KU Leuven, and aPh.D. in Engineering from the faculty of ComputerScience in 2012. He is Assistant Professor at theVrije Universiteit Brussel (VUB), in the Business,Technology and Operations lab of the faculty ofEconomic and Social Sciences & Solvay BusinessSchool. His research lies on the border between datamining and constraint programming, and his maininterest is in integrating domain expertise and userconstraints into data analytics tasks.