You are on page 1of 6

Proceedings of the 2009 IEEE Systems and Information FPM2DMSim.

4
Engineering Design Symposium, University of Virginia,
Charlottesville, VA, USA, April 24, 2009

GPS Tracking of Intermodal Transportation: System Integration


with Delivery Order System
Christopher Heywood, Craig Connor, David Browning, Michael C. Smith, Jianping Wang

Abstract— TMO Global Logistics currently provides cargo from Germany to the central United States. TMO would
tracking services that rely on clients calling the TMO office to then hire the marine, trucking, or rail carriers, obtain the
request estimated time of arrivals (ETAs) and TMO employees required permits, and supervise the cargo’s customs
calling shipping dispatchers to find current cargo locations and processing. During this process, the customer is very
relaying experience-based ETAs to the client. For TMO, anxious about the condition of his valuable cargo, but he is
minimizing the resources (dollars, man-hours, data storage more concerned with its arrival time at its destination. The
space, and bandwidth) spent accepting ETA requests,
client wants to know when he will need resources at the
calculating ETAs, and relaying results can free those resources
for use on other tasks. At the same time, if TMO can provide delivery point to unload and utilize the cargo. When a
its clients with more accurate ETAs retrievable in less time, request for information is received, a TMO employee
relative to competitors, the firm stands to gain significant contacts the carrier currently in possession of the cargo to
market share as a result of these attractive services. obtain its real-time location. The TMO employee then
The team considered issues such as accuracy of the data, ease makes an educated guess, based on his or her experience,
of implementation, resource usage, and client privacy of the and provides the client with an Estimated Time of Arrival
solution. The recommended solution analyzes a particular (ETA). This process consumes TMO personnel resources
route’s historical shipment data and segments each trip into and often results in inaccurate ETAs because TMO must
legs that might be used, in conjunction with legs from other
rely on the carrier to report the location of the shipment.
shipments also stored for future historical analysis, to
reconstruct hypothetical trips and calculate a more accurate TMO engaged a Systems Engineering team from the
ETA than simply using initial and terminal points would University of Virginia to develop a solution to the problem.
permit. To further increase the reliability of the calculation, The company hoped that the team could decrease or
legs with similar start- and endpoints (within a critical radius eliminate the daily time spent by TMO agents on cargo
of the actual leg’s start- and/or endpoints) are also used in the tracking. TMO also wanted greater client satisfaction to
calculation. retain and improve market share, and it decided that TMO
The project team anticipates integrating this ETA approach clients needed a more accurate, reliable ETA. This paper
into TMO Global Logistics’ cargo tracking system. The focuses on the complex problem of generating an accurate
revamped system includes a TMO client interface that maps
ETA for TMO’s clients. The solution proposed here is a
the client’s cargo for added customer satisfaction and a TMO
administrator interface to initialize, and if necessary, manually significant improvement on the existing ETA calculation
adjust shipments. process’s accuracy, and it demands less from TMO
employees.

I. INTRODUCTION II. PREVIOUS WORK

T MO Global Logistics, a Charlottesville-based firm,


provides third-party shipping logistics solutions to its
clients. A client might want to move a load of lumber
There are a number of methods that already exist to
calculate ETA. Mass transit systems, aviation systems, train
systems, and many other modes of transportation for people
and goods benefit from estimates of the arrival time at the
destination. One method for generating an ETA for an
Manuscript received April 6, 2009. This work was supported in part by aircraft involves locating the plane relative to three or four
TMO Global Logistics of Charlottesville, VA.
C. Heywood was a student with the Department of Systems and fixed positions. As the aircraft’s speed is known, the time
Information Engineering, University of Virginia, Charlottesville, VA until it arrives is simply the distance from its current
22904, USA (phone: 301-367-9091; email: cjh3n@virginia.edu). location to its destination divided by its speed [1]. This
C. Connor was a student with the Department of Systems and solution could be applied to any system where the current
Information Engineering, University of Virginia, Charlottesville, VA
22904, USA (email: cmc5ma@virginia.edu). location, the destination location, and the current speed were
D. Browning was a student with the Department of Systems and known. The resulting ETA, however, is most accurate for
Information Engineering, University of Virginia, Charlottesville, VA straight-line, constant-speed modes of transportation like
22904, USA (email: dab4w@virginia.edu). planes or ships and is not sufficient for variable-speed and
M. C. Smith is an Associate Professor with the Department of Systems
and Information Engineering, University of Virginia, Charlottesville, VA route-based modes like trucks and trains.
22904, USA (email: mcs5f@virginia.edu). Another group predicts ETAs for London buses by
J. Wang is a Senior Research Scientist with the Department of Systems determining which two stations the bus is between with the
and Information Engineering, University of Virginia, Charlottesville, VA help of an on-board GPS system. The system then uses the
22904, USA (email: jw9bn@virginia.edu).
known distances between stops to predict an ETA [2]. This

1-4244-4532-5/©2009 IEEE 191


method works well for a predicted route with many stops mode of transportation, and cargo speeds are rarely constant
and a constant low speed of advance. However, it is not or predictable. Finally, an ETA algorithm should be
robust enough to account for new routes, and it would not feasible, and TMO should be able to implement it with
generate a different (and more accurate) ETA whether the currently available resources, including data storage and
bus was closer to the next stop or closer to the stop it just location information from cargo carriers.
left.
ETAs cannot simply be calculated from point A to point B. IV. IMPLEMENTATION
Many trips, including most of TMO’s shipments, travel from
A to B to C, and possibly beyond, with at least one transfer A. General Solution
point (here at B). A transfer point could be a warehouse, a The proposed solution divides a shipment into a series of
change in the mode of transportation, a customs inspection, legs connecting the departure point, intermediary transfer
or any other standardized delay. One study examines the points, and the destination. A TMO agent enters the
process of calculating ETAs from these transfer points, as locations of all points as individual addresses or as
well as estimating time spent at each transfer point. The coordinate points using latitude and longitude. The system
study notes that if an unusual delay occurs at one transfer then tracks the cargo as it transits from point to point as long
point, delays are likely at future transfer points due to a as it receives location information from the cargo unit.
disruption in the planned schedule [3]. Every received location point is stored in a database table
Other researchers suggest queuing models for ETAs. If called HistoricalLeg, along with the mode of transportation,
traffic is very light or nonexistent, travel times for cars on the location of the next transfer point (or, if there are no
roads can be represented with a random variable distribution. more transfer points, the final destination), and two empty
If there is some traffic, it is accurate to calculate ETAs by fields for the arrival time and the elapsed time. These empty
saying that the vehicles are waiting in queues whose fields are filled in when the cargo arrives at the next transfer
attributes can be represented with random variable point. Each HistoricalLeg row thus represents one trip from
distributions [4]. This method requires a lot of data to the point at which location data were received to the next
generate distributions, unfortunately. transfer point.
A study undertaken for the Washington State Department
of Transportation uses historical data between points to
create an ETA algorithm for a mass transit system. This
solution also includes a noise function to account for
variability in travel time: if the first part of a leg takes one
and a half times as long as usual, the system predicts that the
second half of the leg will also take one and a half times as
long as usual [5]. This solution is excellent for short legs
and small networks, but the traffic noise function cannot be
applied across long distances. This study [3] also uses
historical data to estimate leg durations, but its lack of a
method to calculate ETAs when in the middle of a leg
Fig. 1. HistoricalLeg rows generated as the cargo proceeds from A to
recommends it for short journeys as well. B. Sent location information is indicated by triangles. Each dotted
line over the route indicates a startpoint, endpoint, and elapsed time
III. SPECIFICATIONS added to HistoricalLeg.

The system needs to meet a number of requirements in Over time, the HistoricalLeg table grows until the
order to achieve the goals described in the introduction. An commonly traveled path from a given point A to point B is
improved ETA calculation would increase the resolution of littered with start points for HistoricalLeg rows (Fig. 1). If a
the prediction—perhaps the likely hour of arrival, rather than TMO client requests an ETA on in-transit cargo traveling
the day—and would increase the accuracy of the prediction. currently at point P in between the A and B points, the
A more accurate prediction would have a smaller error, or algorithm will query the HistoricalLeg table for the n rows
the difference between the predicted ETA and the actual with the same mode of transportation, the same next transfer
arrival time. These improvements would gratify existing point B, and the same current location P. The system will
TMO clients and hopefully attract new ones. then use the mean of the elapsed time for each row ti to
The ETA calculation should be automated and require calculate the estimated time from the current location to B,
very little input or time commitment from a TMO employee. tPB.
While a completely autonomous system might not be n
plausible at this time, the hour a day TMO agents currently ∑ t
i =1 i
spend on ETA calculation must be reduced.
tPB = (1)
n
The ETA calculation must be flexible and robust. It needs If necessary, it will also calculate the estimated time on
to operate across all modes of transportation, and it should the remaining j complete legs (from B to C, and so on, until
be able to self-correct. TMO’s clients will not accept large destination Z) as well as the estimated time at the remaining
inaccuracies, yet routes can change, cargo can transit on any j transfer points
€ between P and D.

1-4244-4532-5/©2009 IEEE 192


t BZ = t BC + tCD + ...+ tYZ (2) be calculated using Google Maps’ estimated driving time
ttransfer = tB + tC + ...+ tZ (3) function. Estimates for legs that have no Google Maps
estimate (e.g. rail or water segments) would be calculated
ETA = tPB + tBZ + ttransfer (4) using the straight-line distance remaining divided by the
This method allows for a robust ETA calculation (4) at estimated speed of the cargo. The estimated speed would be
€point in the route, as opposed to a manipulation of an
any calculated using the distance covered over the last three data
ETA€only available between points A and B. As certain points divided by the time in which they were covered. A
routes are traveled repeatedly, more data will accumulate TMO agent would approve all times under this method to

and the predicted trip time at any point will become more account for variation in projected times specific to certain
accurate. In addition, most trips will have multiple segments modes of transport (e.g. wide load, hazardous materials).
and transfer points. Some segments will yield rich GPS This method will be built into the system as a backup in
data, but others (such as train segments) may yield sparse case the cargo is in a location outside the sensitivity radii of
location data. These legs without much location data will the startpoints of all legs in HistoricalLegs. However, it is
lead to error in ETA calculations. However, the error will highly inaccurate and will only be used when absolutely
only affect the time calculation for that segment. The ETA necessary. One of these stopgap solutions to an incomplete
will be corrected when the segment is completed and the HistoricalData table will be used temporarily, upon
error will not compound over time. initiation of the system, and will not significantly affect
B. Specific Issues and Solutions ETAs once the system is fully operational.
Unusually high traffic volumes can present another issue:
This solution, while fairly simple, presents complications weighting one trip’s elapsed time too generously. For
that require resolutions. First, except at transfer points, example, consider a trucking leg from A to B that typically
current-location coordinates received will rarely match takes one hour. During that hour, the system receives a
points in the HistoricalLeg table. The less rows the system location on the cargo every 5 minutes, thus placing evenly
can base its average on, the more inaccurate the estimate will spaced entries into the HistoricalLeg table. However, if the
be. Thus, the system uses a “sensitivity radius” r around the truck encounters traffic and stops moving for 15 minutes,
current location P of the cargo to find HistoricalLeg rows three legs would be recorded in the HistoricalLeg table for
that begin near the cargo’s current location. this journey, all of which begin in the same location. The
r = 0.05• AB (5) next time a client’s cargo is near that point, and the client
These rows’ elapsed times are then averaged to generate wants an ETA, the ETA will be calculated using the average
the ETA at the next point. Fig. 2 demonstrates a sensitivity duration of all the legs that start within r of the current
radius and four HistoricalLeg rows that are captured with location. This average will include three legs from the high-
startpoints
€near P and endpoints of B. traffic trip. Thus, the time intensive trip will be weighted
three times as much as any other historical trip.

Fig. 2. Sensitivity radius capturing HistoricalLeg rows with


startpoints near P. Four legs are deemed within r of P and also end at
B. The ETA will be calculated using the average of the elapsed time
of these previous shipments.

The solution will be implemented over time as historical


data are collected to provide a more accurate ETA. When
the system is first integrated into TMO’s existing customer
service operations, there will be no historical data available
for any route. Two proposed solutions will be left to the
discretion of TMO. The website and historical data
components can operate in “stealth” mode, recording
positions and trip lengths without allowing any TMO clients
to use the site. During this phase, TMO agents would
continue to provide ETAs as they do now. After historical
data are accumulated to allow the system to be operational,
clients would be given accounts to the website. The second
option calls for immediate implementation of the system,
including full customer use. Under this option, ETAs would

1-4244-4532-5/©2009 IEEE 193


Fig. 4. Transfer point arrival and departure sensitivity circles. Note
how as the cargo breaches the inner circle, the system considers it to
have arrived at B. However, even though it leaves the arrival circle
before it actually reaches B, it does not leave the departure circle.
Fig. 3. High traffic problem and solution. Assuming constant location Thus the system correctly infers that the cargo has left B when the
data input, in the steady traffic case, only four HistoricalLeg rows are cargo leaves the outer circle.
generated.. However, in the heavy traffic case, six are generated, and
Thus the system incorporates two circles: an arrival circle
three are very close to each other (those rows created when the cargo
encountered the traffic). The corrected heavy traffic case throws out of radius r, and a departure circle of radius 2r, as shown in
all but one of the rows whose origins are too close, thereby weighting Fig. 4. The cargo is considered to have arrived at the
the high traffic leg fairly in the HistoricalLeg table and in the next transfer point, and thus started the clock, when the inner
ETA calculation.
sensitivity circle is breached. The cargo then is considered
This concern is corrected by once again utilizing to have departed the transfer point when it leaves the outer
sensitivity radii. A row will only be entered in HistoricalLeg circle, thereby setting tB to zero and beginning the leg to C.
if the current location is at least 2r away from the startpoint
of the last leg entered in the table due to the same shipment. C. Testing
This solution avoids including the same trip multiple times
in an average when creating a future ETA. No sensitivity
radius will contain more than one leg from any previous
individual shipment, thus avoiding inaccurate weighting of
traffic-heavy shipments. This is summarized in Fig. 3.
Transfer points provide another concern for an ETA
calculation. If cargo is traveling from A to B to C, and B is
projected to be a one hour transfer point, the ETA will be the
sum of the projected time from B to C, the one hour transfer
time at B, and the estimated time from the current location P
to B. As soon as the cargo arrives within one sensitivity
radius of B, it is assumed to be at B, and thus tPB is
considered zero. A clock tB begins to count down from one
hour, representing the projected time left at B. However, the Fig. 5. Test path. Cargo at P has not yet traversed the dotted segment.
transfer estimate is never perfectly accurate, and so when the In order to show the importance of depth of data on the
vehicle leaves the vicinity of transfer point B, tB is set to usefulness of the historical analysis estimated time of arrival,
zero, and the final ETA equals tPC. If a truck driver, the team needed to compare the accuracy of the estimations
however, enters and leaves the sensitivity circle around B with the frequency of leg start- and endpoints in the area. In
before he actually arrives at the transfer point (a feasible the absence of reliable real trip data, the team simulated data
situation), the system might consider the transfer point for a hypothetical highway. The highway is 1000 miles long
finished and underestimate the time remaining before and has a speed limit of 65 mph. The estimated time of
arrival. arrival is requested when the driver is at the 500 mile marker
(shown with a P in fig. 5). The time taken to traverse any
stretch of the hypothetical highway was calculated by
assuming the driver would drive at the speed limit, and then
adding on a delay represented by a time pulled randomly
from a normal distribution with a mean of two hours and
standard deviation of 50 minutes. The team realized that

1-4244-4532-5/©2009 IEEE 194


these traffic conditions do not exist exactly in the real world, order to be initialized by a technician. The system
but also thought that simulation was better than inaccurate automates the rest of the process by storing historical data
data. The system looked for historical legs with start- and and automatically calculating ETA. It also accepts data of
endpoints within 25 mile radii (5%) of the current location any frequency and from any source, as long as it arrives in
and destination (shown with a B in fig. 5), and used those to address or latitude/longitude format, allowing it to be
calculate an estimated time of arrival. The team then multimodal and flexible.
compared this estimation with the actual travel time, which The system also allows for future analysis of its accuracy
was drawn from the same normal distribution as the legs and of possible trends among errors. Every calculated ETA
were, and was replaced for each comparison. The team ran is stored in a table called EtaAnalysis, where the predicted
the experiment with 20 leg startpoints dropped randomly arrival date and time are recorded next to the actual arrival
between the start of the 1000-mile stretch of highway date and time. This allows for future enhancements of the
(shown with an A in fig. 5) and the end of the stretch of algorithm, including an examination of trends over time and
highway (marked by the T), and then repeated the periodic or cyclical variations in the error due to time of
experiment with 50 points (0.05 drops per mile), 100 points day, week, or year.
(0.1 drops per mile), 300 points (0.3 drops per mile), and
600 points (0.6 drops per mile). V. CONCLUSIONS
The results showed that the historical analysis ETAs get TMO hopes to increase market share and maximize the
more robust as the amount of data increases. This happens
utility of its resources. An automated, more accurate ETA
for two reasons. First, the probability of being in an area
can contribute to the solution of both these problems. When
with no historical leg startpoints within the sensitivity radius
the historical analysis ETA is combined with a map-based
decreases as the number of GPS inputs received along that
stretch of road increases. This relationship is shown in fig. user interface (fig. 8), TMO clients will have a powerful
6. Secondly, the accuracy of the estimations also increases, new tool to monitor their shipments. The system
although very slightly, as shown in fig. 7. architecture (fig. 9) remains hidden, but accurate and
automatic ETAs are outputted. Costs and stresses due to
uncertain cargo arrival times will be significantly reduced.

Fig. 6. More drops/mile leads to historical data on a greater


percentage of the road in simulation.

Fig. 8. TMO client interface conception.

Fig. 7. Accuracy of ETA in simulation.


The proposed system thus meets the requirements of
increased resolution and accuracy. It also requires less time
commitment from TMO employees, as it only requires each

1-4244-4532-5/©2009 IEEE 195


Fig. 9. Architecture of information flow in the system. Location
information flows from the cargo as it passes through various points
and legs to the data receiver, at which point it enters the designed
system. The location information is then sent to storage, where it is
accessed by the TMO admin interface and the ETA algorithm. The
algorithm stores the estimates in the database and later compares them
to the true values for accuracy. The TMO agent initializes shipments
in the database.. The end result is an accurate ETA estimate sent to
the client interface.

ACKNOWLEDGMENT
The team thanks the employees of TMO Global Logistics
for their support. Joseph “J.C.” Toe and James Holland
were especially helpful as the team struggled with problem
comprehension and needed data to analyze.

REFERENCES
[1] B. S. Levy and S. Bedada, “A real-time ETA-to-threshold prediction
tool,” 2006 IEEE/AIAA Digital Avionics Systems Conference, pp. 1-
12.
[2] O. Rashid, P. Coulton, R. Edwards, A. Fisher, and R. Thompson,
“Mobile information systems providing estimated time of arrival for
public transport users,” Spring 2005 IEEE Vehicular Technology
Conference, vol. 5, pp. 2765-2769.
[3] B. Kara and B. Tansel, “The latest arrival hub location problem,” in
Management Science, vol. 47, no. 10, October 2001, pp. 1408-1420.
[4] A. Miller, “A queueing model for road traffic flow,” in Journal of the
Royal Statistical Society, Series B (Methodological), vol. 23, no. 1,
1961, pp. 64-90.
[5] Z. Wall and D. J. Dailey, “An algorithm for predicting the arrival time
of mass transit vehicles using automatic vehicle location data,”
Transportation Research Board, Washington, D. C., Jan. 10-14, 1999.

1-4244-4532-5/©2009 IEEE 196

You might also like