Large- ScaLe OppOrtuniStic SenSing

estimating OriginDestination Flows using Mobile phone Location Data

Using an algorithm to analyze opportunistically collected mobile phone location data, the authors estimate weekday and weekend travel patterns of a large metropolitan area with high accuracy.

rigin-destination (OD) matrices represent one of the most important sources of information used in the strategic planning and management of transportation networks. A precise calculation of OD matrices is an essential component in enabling administrative authorities to optimize the use of their transportation networks, not only to benefit users on their Francesco Calabrese daily journeys, but also to and Giusy Di Lorenzo plan investments required to IBM Dublin Research Laboratory adapt these infrastructures to envisaged future needs. TraLiang Liu and Carlo Ratti ditionally, urban planning and Massachusetts Institute transportation engineering of Technology use household questionnaires or census and road surveys to develop methodologies for OD matrix estimation. This approach has two main drawbacks: calculating an OD matrix, from the initial data gathering to the exploitation of the first results, can take years and produce only a snapshot of the travel demand; and the collected data has shortcomings in terms of both spatial and temporal scale. Sensor-based OD estimation methods use street sensors such as loop detectors and video

cameras together with traffic-assignment models. Analogous methods have been developed using probe vehicles, in which vehicle traces serve as data sources.1,2 Those methods are, however, limited by the fact that models are often underdetermined because the number of parameters to be estimated is typically larger than the number of monitored network links.3 On the other hand, the wide deployment of pervasive computing devices (mobile phone, smart cards, GPS devices, digital cameras, and so on) provides unprecedented digital footprints, telling where people are and when theyre there. Earlier projects have used different methodologies for detecting the presence and movement of crowds through their digital footprints (Flickr photo, mobile phone logs, smart card records, and taxi/bus GPS traces).46 This fine-grained analysis could dramatically increase our understanding of the use of space and daily commuting flows for urban mobility planning and management. Thus, its no surprise that the idea of using mobile phones to monitor traffic conditions isnt new, as we discuss in the Related Work in Analyzing Traffic Flow sidebar. Although the results from these other studies show great potential for using cellular probe trajectory information to estimate travel demand, all methods must overcome several shortcomings before they can be put into practice. Indeed, as

36 PER VA SI V E computing

Published by the IEEE CS n 1536-1268/11/$26.00 2011 IEEE

related Work in analyzing traffic Flow

everal related studies on analyzing traffic flow have been published in recent years. Raffaele Bolla and Franco Davoli presented a model for estimating traffic using an algorithm that calculates traffic parameters on the basis of mobile phone location data.1 Researchers in Rome developed a case study for real-time urban monitoring using aggregated mobile phone data to monitor traffic and movement of vehicles and pedestrians.2 Randal Cayford and Tigran Johnson analyzed the main parameters to be considered, namely precision, metering frequency, and the number of localizations necessary to achieve accurate traffic descriptions.3 Several companies worldwide, including ITIS Holdings (Britain), Delcan (Canada), CellInt (Israel), and AirSage and IntelliOne (USA), have begun developing commercial applications of mobile phone-based traffic monitoring. With the specific goal of measuring OD flows, different mobile phone signaling datasets have been considered and simulated to evaluate the feasibility of estimating trips. Initial work used billing data, consisting of cell phone tower information every time a phone received or made a call.4 Other research has used mobile phone positions every two hours to infer trips, 5 location updates to infer mobile phone movement,6 and cell phone tower handover information.7 A recent effort estimated the daily OD demand using simulated cellular probe trajectory information (extracted from location updates, handover, and transition of timing

advance values) and tested the methodology via the VisSim simulation.8

REfEREnCES
1. R. Bolla and F. Davoli, Road Traffic Estimation from Location Tracking Data in the Mobile Cellular Network, Proc. IEEE Wireless Comm. and Networking Conf., vol. 3, IEEE Press, 2000, pp. 11071112. 2. F. Calabrese et al., Real-Time Urban Monitoring Using Cell Phones: A Case Study in Rome, IEEE Trans. Intelligent Transportation Systems, vol. 12, no. 1, 2011, pp. 141151. 3. R. Cayford and T. Johnson, Operational Parameters Affecting Use of Anonymous Cellphone Tracking for Generating Traffic Information, Proc. Transportation Research Board Ann. Meeting, 2003. 4. J. White and I. Wells, Extracting Origin Destination Information from Mobile Phone Data, International Conference on Road Transportation and Control, IEE, 2002, pp. 3034. 5. C. Pan et al., Cellular-Based Data-Extracting Method for Trip Distribution, J. Transportation Research Board, vol. 1945, 2006, pp. 3339. 6. N. Caceres, J. Wideberg, and F. Benitez, Deriving Origin Destination Data from a Mobile Phone Network, Intelligent Transport Systems, vol. 1, no. 1, 2007, pp. 15 26. 7. K. Sohn and D. Kim, Dynamic Origin-Destination Flow Estimation Using Cellular Communication System, IEEE Trans. Vehicular Technology, vol. 57, no. 5, 2008, pp. 2703 2713. 8. Y. Zhang et al., Daily O-D Matrix Estimation Using Cellular Probe Data, Proc. Transportation Research Board Ann. Meeting, 2010.

Yi Zhang and his colleagues note,7 field tests are needed for the following reasons: real coverage areas of cell phone towers are quite different from simulated ones, and vary from urban to rural areas; validation of methods to determine a trips origin and destination should be performed using real individual mobility data; real mobility and calling patterns should be included in the analysis because they crucially influence the methods performance; existing OD matrices should be used as ground truth to verify the correctness of the estimated results. Our methodology uses opportunistically collected mobile phone location

data to estimate dynamic OD matrices. We address concerns using a real mobility and calling dataset from 1 million mobile phone users. We use the Boston metropolitan area as a case study and validate our methodology using census survey data for both county and censustract levels.8 To our knowledge, both the methodology developed and the data precision and amount are unique.

when the user connects to the Internet (for example, to browse the Web, or through email programs that periodically check the mail server). In this article, we refer to these events as network connections. The events represent a superset of those in the call detail records used elsewhere.9,10 We analyzed 829 million mobile location data for 1 million devices collected by AirSage (www.airsage. com). This data included not only the ID of the cell tower the mobile phone was connected to, but also an estimation of its position within the cell, which is generated through triangulation by AirSages Wireless Signal Extraction technology. Each location measurement m i M is characterized by a position pmi expressed

Mobile phone Dataset

The considered dataset consists of anonymous location measurements generated each time a device connects to the cellular network, including: when a call is placed or received (both at the beginning and end of a call); when a short message is sent or received;

OCTOBERDECEMBER 2011

PER VA SI V E computing

37

Large-ScaLe OppOrtuniStic SenSing

0.10 0.09 0.08 0.07 Distribution 0.06 0.05 0.04 0.03 0.02 0.01 0 102 101 100 101 102 Interevent time (minutes) 103 104 105 First quantile Median Third quantile

To alleviate the effects of localization errors and event-driven location measurements on individual trip determination, we apply a low-pass filter with a 10-minute resampling rate to the raw data. This follows an approach tested with data from Rome, Italy.11 In addition, because fewer localization errors might still generate fictitious trips, we adapt a preprocessing step used to analyze GPS traces. This step uses clustering to identify minor oscillations around a common location. The approach used to handle location errors and identify meaningful locations in a users travel history can be understood as follows: We begin with a measurement series Ms = {mq, mq+1, , mz} Mz-q-1, q > z, derived from a series of network connections over a certain time interval T = t mz t mq > 0. We define an area with radius S (in this case, 1 km to account for the localization errors estimated by AirSage), such that max distance ( pmi , pmj ) < S q i, j z.

Figure 1. Characterization of individual calling activity for the entire population, in terms of time between two network connection events. Graphs show the distributions of the median (solid line), first quantile (dash-dotted line), and third quantile (dashed line) of individual interevent time.

in latitude and longitude and a timestamp t mi . To infer trips from these measurements, we first characterized the individual calling activity and verified that its frequent enough to allow monitoring the users movement over time with a fine enough resolution. For each user, we measured the interevent timethat is, the time interval between two consecutive network connections (similar to what Marta Gonzlez and her colleagues measured10). The average interevent time measured for the entire population was 260 minutes, much lower than Gonzlez and her colleagues measurement (500 minutes) because were also considering mobile Internet connections. Because the distribution of interevent times for a user spans several temporal scales, we further characterized each calling activity distribution by its first and third quantile and the median. Figure 1 shows the distribution of the first and third quantile and

the median for all available users into the dataset. The arithmetic average of the medians is 84 minutes (the geometric average of the medians is 10.3 minutes) with results small enough to detect changes of location where the user stops for as little as 1.5 hours. Mobile-phone-derived location data has lower resolution than GPS data. Internal and independent testing suggest an average uncertainty radius of 320 meters, and a median of 220 meters. Moreover, at some peak usage periods, additional location errors can be introduced when users are automatically transferred by the network from the closest cellular tower to one thats further away but less heavily loaded.

All consecutive points pj M s for which this condition holds can be fused together such that the centroid becomes a virtual location ps = (z q)1

i = q pm (the centroid
i=z
i

of the points),

that is a trips origin or destination. Once the virtual locations are detected, we can evaluate the stops (virtual locations) and trips as paths between users positions at consecutive virtual locations. Each trip trip(u, o, d, t) is characterized by user ID u, origin location o, destination location d, and starting time t. Once trips are extracted, we use the following procedure to derive OD flows: The geographical area under analysis is divided into regions: regioni , i = 1, , n.

Origin-Destination estimation Method

The procedure for estimating dynamic OD matrices consists of two steps: trip determination and origin-destination estimation.

38

PER VA SI V E computing

www.computer.org/pervasive

100

100 101 102

Weekday Weekend

101

Distribution

Distribution 0.5 1.0 1.5 Trip length (log10 km) 2.0 2.5

102

103 104 105 106

103

104

0.5

(a)

(b)

0 0.5 1.0 Number of trips (log10)

1.5

2.0

Figure 2. Statistics on the trips detected from the mobile phone location data: (a) trip-length distribution (curve interpolated with P(x) = (x + 14.6) -0.78exp(-x/60) with R 2 = 0.98), and (b) number of individual trips per day distribution.

Origin and destination regions, together with starting time, are extracted for each trip of each user trip(u, o, d, t). Trips with the same origin and destination regions are grouped together at different temporal windows tw, for example, weekly, daily, and hourly: m(i, j, tw) =
o regioni , d region j , t tw

trip(u, o, d , t).

The result is a 3D matrix M 3 whose element m(i, j, tw) represents the number of trips from origin region i to destination region j starting within the time window tw.

case Study and comparison

Based on the area covered by the mobile phone locations dataset, we analyzed movement among areas in eight counties in eastern Massachusetts (Middlesex, Suffolk, Essex, Worcester, Norfolk, Bristol, Plymouth, and Barnstable) with an approximate population of 5.5 million people. To simplify the analysis, we randomly selected 25 percent of the available users and extracted traces for them.

trip characterization As a first analysis, we studied the triplength distribution (see Figure 2a), showing that trips range from 1 to 300 km. We determined the trip length x by calculating the Euclidean distances between the trips origins and destinations. The distribution is well approximated by P(x) = (x + 14.6) -0.78 exp(-x/60) with R 2 = 0.98, which confirms previous findings.10 The slightly different coefficients found in this case could be because of the different build environments in Europe and the US.12 To check the plausibility of our segmentation of the trajectory in trips, we compute the same statistics computed on the number of individual trips per day. Figure 2b shows the distribution over the entire population, separating weekday and weekend trips. We obtain an average of 5 trips per day during the week, and 4.5 on weekends. This number is reasonable when compared to the US National Household Travel Survey (see http://nhts.ornl.gov), which evaluated this number to be between 4.18 during the week and 3.86 on weekends. (Reasons for these differences include several years of difference between when the two datasets were collected, and the

fact that NHTS is based on a sample over the entire US population, so not focused on the behavior of people in the Boston metropolitan area.) To evaluate whether we have sampling biases in our data, we computed the home location distribution estimated from the mobile phone data, and compared it with data from the US 2000 Census. To detect a home location, we first group together geographic regions that are close in space, creating a grid in space where the side of every cell is 500 meters long. For each cell, we evaluate the number of nights the user connects to the network in the nighttime interval while in that cell, and select as a home location the cell with the greatest value. (We used 6 p.m. to 8 a.m. as the nighttime interval based on statistics from the American Time Use Survey; www. bls.gov/tus/charts/work.htm.) To validate the home location distribution, we compared it with population data from the US 2000 Census, at the census-tract level.13 Within the eight selected counties are 1,171 distinct census tracts, with populations ranging from 70 to 12,000 people (on average 4,705), and an area ranging from 0.08 to 203 km2 (on average 10.8 km2).

OCTOBERDECEMBER 2011

PER VA SI V E computing

39

Large-ScaLe OppOrtuniStic SenSing

Cell phone user density (person/sq mi)

0.000000000122.6 122.7287.3 287.4521.8 521.9847.9 848.01222

12231653 16542158 21592888 28894353 43547267

(a)
0

1
Estimated tract population density (log10)

The tracttract worker flows data shows the number of workers in each tract of work by tract of residence. Workers are defined as people 16 years old and older who were employed and at work, full or part time, during the census reference week (generally the last week of March). The data contains the number of workers in the flow who were allocated to tract, place, and county of work. Given the two levels of granularity (tract and county) available in the CTPP dataset, we computed our OD estimates at two levels of aggregation. Because commuting flow generally accounts for two trips (home to work and work to home), we considered undirected flows between two locations to compare our OD estimations. For each granularity, we computed the average daily number of trips: mAll (i, j) = KAll (m (i, j, tw) + m ( j, i, tw)), # days tw = day

i = 1,..., n

j = 1,..., i 1

7 5

4.5

(b)

4.0 3.5 3.0 2.5 2.0 Census tract population density (log10)

1.5

1.0

Figure 3. US 2000 census tract population density and cell phone users estimated home locations density. (a) Census tract population density derived from cell phone users estimated home locations. (b) To compare these two densities, we multiplied cell phone-based population density by 100/4.3 to account for the percentage of the population being monitored. Error bars represent the standard error.

where K All is a scaling factor we use to compare our estimation with the census estimations. Moreover, because according to the definition, the census dataset includes only commuting trips, we evaluated the average daily number of trips made only on weekday mornings (6 to 10 a.m.) from the estimated home to estimated work location (estimated as the most frequent stop area on weekday mornings between 8 and 10 a.m.):
mWM (i, j) = KWM (m(i, j, tw) + m( j, i, tw)), # weekdays tw = wm

The census tract population estimated using cell phone users home locations scales linearly with the census population, as Figure 3 shows, corresponding to an average 4.3 percent of the population being monitored. OD Flows characterization To validate the accuracy of the OD matrices produced using the mobile

phone traces, we used the most recent tracttract worker flows dataset from the Census Transportation Planning Package. 8 CTPP is a tabulation of responses from households completing the census long form. It is the only census product that summarizes data by place of work and tabulates the flow of workers between home and work.

i = 1,..., n

j = 1,..., i 1

where KWM is a scaling factor. Finally, we also compared our predictions with the well-known and widely used gravity model:14
mGravity (i, j) = KG Pi Pj di2 j , , i = 1,..., n j = 1,..., i 1, ,

40

PER VA SI V E computing

www.computer.org/pervasive

6 5 4 3 2 1 0
All trips Weekday mornings trips Gravity model

4.0 3.5 Estimated number of trips (log10) 5 6 3.0 2.5 2.0 1.5 1.0 0.5 0 0.5 0 0.5 1.0 1.5 2.0 2.5 Census number of trips (log10) 3.0 3.5

Estimated number of trips (log10)

(a)

2 3 4 Census number of trips (log10)

(b)

Figure 4. Comparison between mobile phone and census origin-destination (OD) estimates. Error bars show one standard deviation from the average: (a) county level showing all trips, weekday morning trips, and the gravity model; and (b) tract level showing weekday morning trips.

where KG is a scaling factor, and di,j is the Euclidean distance (in kilometers) between the centroids of the regions. Figure 4a shows the county-level results. The plots correspond to models that minimize the least square errors, using K All = 16.9 for the prediction made with the average number of trips in a day mAll, KWM = 71.4 for the prediction made with the average number of trips on weekday mornings mWM , and KG = 58.4 for the gravity model mGravity. Correlations show encouraging results, with R 2 = 0.59 for the gravity model, R 2 = 0.73 for the prediction made with all trips, and the best result R 2 = 0.76 for predictions made with weekday morning trips only. The resulting high correlation shows that the estimated OD matrices resemble OD matrices generated using completely different information. Using the best model mWM , we compared our results with the tract-level census data. At this level, noise is more evident (see Figure 4b), but we can still see on average a good linear relationship between census estimation and our estimationR 2 = 0.36 in this case, which is high compared to the R2 = 0.10

of the gravity model. We also evaluated more sophisticated gravity-like models by optimizing the d exponent and substituting the populations with the total estimated number of trips outgoing or incoming for an area, but still obtained R 2 < 0.3. The relatively low value of R 2 compared to the county-level analysis is partially because the relationship seems less linear when the census estimates fewer than 10 trips from tract to tract. This might be because census flows are estimated from a subsample, which might result in small numbers for particular pairs of census tracts. Moreover, census estimates werent available for the same year as the mobile phone data, and trip origins and destinations might have slightly changed (at this high level of spatial detail) between the two monitored periods. We note that the scaling factor KWM used for the last model mWM corresponds to a share of monitored trips thats about 1.4 percent compared to the census estimations. This factor can be explained by the percentage of mobile phones selected (about 4.3 percent) and by the calling activity, which isnt very high in the morning.

Other elements, such as the fact that were monitoring not only commuting flows, might explain the remaining difference. Estimating KWM lets us extrapolate the ODs computed using the mobile phone data to the whole population.

new potential
We see a great deal of potential for this approach. We should note, though, that OD flows data estimated through census surveys have the following limitations:8 The decennial census monitors usual days to avoid local or regional anomalies, such as transit strikes or severe weather, on a single sampling day. This tends to hide the less common uses, such as individuals telecommuting once every two weeks or carpooling once a week because of ever-changing life and work patterns. According to the definition, the census dataset doesnt include nonwork trips, and modelers must develop relationships between work and nonwork trips. T he c ensu s d at a is based on a fixed-point snapshot approach, so

OCTOBERDECEMBER 2011

PER VA SI V E computing

41

Large-ScaLe OppOrtuniStic SenSing

2.1 2.0 1.9 1.8 1.7 1.6 1.5

106

15

104
Columbus Day Other Mondays

Number of trips

Number of trips

10

Sun Mon Tue Wed Thu Fri Sat Sun Mon Columbus Day Tue Wed Thu Fri Sat Sun Mon Tue Wed Thu Fri Sat

(a)

(b)

10 Hour

15

20

Figure 5. Temporal variation in the number of trips detected from the mobile phone location data: (a) number of trips per day, over a three-week interval; and (b) number of trips per hour, over three different Mondays.

transportation planners can only interpret data over geographic space, rather than over time. Compared with traditional census data, however, our methodology to detect OD matrices from mobile phone traces has several advantages: It can capture the weekday and weekend patterns as well as seasonal variations. It can capture work and nonwork trips, which is essential for trip chaining and activity-based modeling. Thus, these matrices could then complement traditionally generated OD matrices, providing a fine-grained spatiotemporal pattern of mobility. temporal analysis Whereas the census gives only static information about OD flows, the OD matrices derived from mobile phone data let us appreciate the differences in travel demand over time. Figure 5a shows the total daily travel demand for three weeks in October 2009. A weekly pattern clearly appears in the travel demand, with the minimum on weekends

(especially Sundays) and the maximum on Fridays. Moreover, Figure 5a shows a change in travel demand on the second Monday (day 9 in the figure), corresponding to Columbus Day. For a better look at this pattern, we plot the hourly travel demand for Columbus Day compared to other Mondays (see Figure 5b). We clearly see a higher travel demand in the first two hours of the day, followed by lower demand from hour 4 to hour 9, and from hour 12 to hour 20, because of the holiday. Spatiotemporal analysis Our methodology can capture finegrained OD matrices in both spatial and temporal scale, essential data for understanding transport demand and transport modeling, especially during special events. For example, Figure 6 compares the incoming flows toward Fenway Park, Bostons baseball stadium. We compare two days: Sunday, 11 October, when the Boston Red Sox played the Los Angeles Angels in a postseason game, and a Sunday with no special events. As the figure shows, we can capture the increasing incoming flow caused by the event, both in terms of new trip origins, and in

flow volume. Further studies with the same dataset have also shown regular spatial patterns of attendee origins based on event type, information that would be valuable for event management.5 s this study shows, pervasive datasets such as mobile phone traces provide rich information to support transportation planning and operation. Meanwhile, some related limitations should also be addressed when applying these datasets in mobility analysis. The localization error, for example, limits the minimum size of the regions that can be considered. Other elements that can affect the statistical results include: the market share of the mobile phone operator from which the dataset is obtained, the potential nonrandomness of the mobile phone users (for example, teenagers), calling plans, which can limit the number of samples acquired at each hour or day, and the number of devices that each person carries.

42

PER VA SI V E computing

www.computer.org/pervasive

(a)

(b)

Figure 6. Incoming trips in the Fenway Park area. Each line connects a detected trip origin area to the Fenway Park: (a) normal Sunday, and (b) day of a Red Sox game. Flow volume is represented by the lines thickness.

the AUThORS
francesco Calabrese is an advisory research staff member at the IBM Dublin Research Laboratory and a research affiliate at the SENSEable City Laboratory of the Massachusetts Institute of Technology. His research interests include ubiquitous computing, intelligent transportation systems, urban network analysis, and distributed control systems design. Calabrese has a PhD in computer and systems engineering from the University of Naples Federico II, Italy. Hes a member of IEEE and the IEEE Control Systems Society. Contact him at fcalabre@ie.ibm.com. Giusy Di Lorenzo is a research scientist at the IBM Dublin Research Laboratory and a research affiliate at the SENSEable City Laboratory of the Massachusetts Institute of Technology. Her research explores the analysis of urban dynamics and human mobility using data gathered from sensor networks. In particular, shes interested in developing context-aware applications to improve urban living experiences of citizens. Di Lorenzo has a PhD in computer and systems engineering from the University of Naples Federico II. Contact her at giusydil@ie.ibm.com. Liang Liu is the deputy director of the Creative Industries Park in Shanghai and Changzhou. He was a postdoctoral associate at the SENSEable City Laboratory of the Massachusetts Institute of Technology. Liu has a PhD in urban planning and management from Tongji University, China. Contact him at liuliang@mit.edu.

Carlo Ratti is an architect, engineer, and associate professor of practice at the Massachusetts Institute of Technology, and director of MITs SENSEable City Laboratory. Ratti has an MSc in civil structural engineering from both the Polytechnic University of Turin and the School of International Management (ENPC) in Paris, and a PhD in architecture from the University of Cambridge. He is a member of the Order of Engineers in Turin and the Architects Registration Board (UK). Contact him at ratti@mit.edu.

Moreover, because the considered dataset is event driven (location measurements available only when the device makes network connections), users connection patterns can affect the possibility of capturing more or fewer trips. This last limitation could be solved by continuous location readings from GPS devices, which would require the users consent. A hybrid approach could integrate both event-driven and

continuous location measurements, as the current method can be easily generalized to different datasets with different spatiotemporal resolutions. Nonetheless, the analysis performed on the interevent time, the spatial distribution of mobile phone users, and comparisons with census estimations confirm that the mobile phone data represent a reasonable proxy for human mobility.

Future work will involve reproducing the analysis for other cities to understand which parameters influence the scaling factors to be used to extrapolate the ODs computed using the mobile phone data to the whole population. The research output will give transportation planners an automatic and systematic way to understand the dynamics of daily mobility in a real complex metropolitan area.

OCTOBERDECEMBER 2011

PER VA SI V E computing

43

Large-ScaLe OppOrtuniStic SenSing

ACknOwlEDGMEnTS
We thank Moshe E. Ben-Akiva for his feedback and AirSage for providing the mobile phone location dataset.

Content, IEEE Pervasive Computing, vol. 7, no. 4, 2008, pp. 3643. 5. F. Calabrese et al., The Geography of Taste: Analyzing Cell-Phone Mobility and Social Events, Proc. Intl Conf. Pervasive Computing, Springer, 2010, pp. 2237. 6. D. Quercia et al., Mobile Phones and Outdoor Advertising: Measurable Advertising, IEEE Pervasive Computing, vol. 10, no. 2, 2011, pp. 2836. 7. Y. Zhang et al., Daily O-D Matrix Estimation Using Cellular Probe Data, Transportation Research Board Ann. Meeting, 2010. 8. CTPP Census Transportation Planning Package, US Department of Transportation, Census Transportation Planning Products, 2010; www. fhwa.dot.gov/planning/census_issues/ ctpp/. 9. J. White and I. Wells, Extracting Origin Destination Information from Mobile Phone Data, International Conference

on Road Transportation and Control, IEE, 2002, pp. 3034. 10. M. Gonzalez, C. Hidalgo, and A.-L. Barabasi, Understanding Individual Human Mobility Patterns, Nature, vol. 453, no. 7196, 2008, pp. 779782. 11. F. Calabrese et al., Real-Time Urban Monitoring Using Cell Phones: A Case Study in Rome, IEEE Trans. Intelligent Transportation Systems, vol. 12, no. 1, 2011, pp. 141151. 12. L. Liu et al., The Law of Inhabitant Travel Distance Distribution, Proc. European Conf. Complex Systems, 2009. 13. MassGIS, Census 2000 Tracts Datalayer Description, 2010; www.mass.gov/ mgis/cen2000_tracts.htm. 14. J. Anderson, A Theoretical Foundation for the Gravity Equation, Am. Economic Rev., vol. 69, no. 1, 1979, pp. 106116.
Selected CS articles and columns are also available for free at http://ComputingNow.computer.org.

REFEREnCES
1. X. Zhou and H. S. Mahmassani, Dynamic Origin-Destination Demand Estimation Using Automatic Vehicle Identification Data, IEEE Trans. Intelligent Transportation Systems, vol. 7, no. 1, 2006, pp. 105114. 2. S. Baek et al., Method for Estimating Population OD Matrix Based on Probe Vehicles, Korean Soc. Civil Engineers (KSCE) J. Civil Eng., vol. 14, no. 2, 2010, pp. 231235. 3. M.L. Hazelton, Some Comments on Origin-Destination Matrix Estimation, Transportation Research Part A: Policy and Practice, vol. 37, no. 10, 2003, pp. 811822. 4. F. Girardin et al., Digital Footprinting: Uncovering Tourists with User-Generated

AdvertiSer informAtion october/december 2011

Advertising Personnel

Marian Anderson: Sr. Advertising Coordinator Email: manderson@computer.org Phone: +1 714 816 2139 | Fax: +1 714 821 4010 Sandy Brown: Sr. Business Development Mgr. Email sbrown@computer.org Phone: +1 714 816 2144 | Fax: +1 714 821 4010
Advertising Sales Representatives (display) Advertising Sales Representatives (Classified Line)

Western US/Pacific/Far East: Eric Kincaid Email: e.kincaid@computer.org Phone: +1 214 673 3742 Fax: +1 888 886 8599 Eastern US/Europe/Middle East: Ann & David Schissler Email: a.schissler@computer.org, d.schissler@computer.org Phone: +1 508 394 4026 Fax: +1 508 394 4926

Greg Barbash Email: g.barbash@computer.org Phone: +1 914 944 0940 Fax: +1 508 394 4926
Advertising Sales Representatives (Jobs Board)

Greg Barbash Email: g.barbash@computer.org Phone: +1 914 944 0940 Fax: +1 508 394 4926

44

PER VA SI V E computing

www.computer.org/pervasive