Integration of Mobile RFID and Inertial Measurement

for Indoor Tracking of Forklifts Moving Containers

M. Fodor

, O. Gusikhin

, E. Tseng

and W. Wang

Ford Motor Company

Research and Innovation Center

20000 Rotunda Drive, Dearborn, MI 48121, U.S.A.

Ford Motor Company

Walton Hills Stamping Plant

7845 Northfield Road, Bedford, OH 44126, U.S.A.

Abstract. Described are the motivation and a method for combining readings

of RFID tags placed in static locations in a plant environment using a mobile

reader attached to a forklift and inertial measurements of the motion of the

forklift to track its position and ultimately the positions of containers it moves

within the plant. Strengths and limitations of RFID-based and inertial-based

methods are presented along with an algorithm for postprocessing and

combining the readings of each. The accuracy of the resulting position

estimates are shown to be on the same order as the accuracy of RFID reader

range variation.

1 Introduction

Desire for real-time visibility of inventory and assets across the supply chain is

driving fast adoption of advanced information technology for location tracking.

Location information eliminates the need for non-value added inventory search

activities and creates a foundation to further optimize the operational efficiency of

business units [1]. While the integration of RFID, GPS, and wireless communication

is already common for the logistics industry to track products in routes between

supply chain sites [2], tracking within indoor environments such as plants and

warehouses is still a challenging problem.

Advancements in RFID technology have facilitated locating tagged assets

indoors [3]. In particular, the field of Real Time Locating Systems (RTLS) is rapidly

growing, primarily employing active RFID. Existing RTLS solutions differ in

operating frequency, methods, granularity, accuracy, and the resulting cost of

infrastructure and operation. Despite significant development, RTLS is presently

prohibitive for commodity item tracking due to the substantial cost of active RFID

and is typically employed only for tracking personnel and expensive resources.

Tracking commodity inventory or containers is usually accomplished using

passive RFID and fixed readers operating at "choke points", providing zoning

location of a given inventory item. For a large area, the infrastructure cost can be

very high, and often fixed readers are installed only at shipping and receiving doors,

providing inventory information but not location. The recent introduction of forklift-

Fodor M., Gusikhin O., Tseng E. and Wang W. (2009).

Integration of Mobile RFID and Inertial Measurement for Indoor Tracking of Forklifts Moving Containers.

In Proceedings of the 3rd International Workshop on Intelligent Vehicle Controls & Intelligent Transportation Systems, pages 120-129

 SciTePress

mounted mobile RFID readers [4] addresses the problem of limited visibility into

inventory locations associated with choke points. In this case the locations of the

inventory items can be recorded from the location of the RFID reader at the point of

unloading. Provided that these items will be moved using only vehicles equipped

with the reader and the location of the reader is known, this method can offer

relatively reliable location records. A number of different approaches have been

proposed for localization of mobile readers. In addition to the RFID-based RTLS

mentioned above [5, 6], other technologies include Wi-Fi, Ultrasonic, and Infrared.

Often these network-centric approaches are hampered by the need for a substantial

initial investment in infrastructure to ensure the required coverage and accuracy. An

alternative approach relies on sensors and instrumentation installed on the delivery

vehicle to track its location.

There is a substantial body of knowledge related to vehicle-centric localization

methods developed within the field of robotics [7] and successfully applied in

industrial settings [8]. The most common approaches are dead reckoning and the use

of landmarks. Dead reckoning is a method of finding the relative position of a mobile

device from a previous known position using inertial measurements or odometry.

Challenges with this method include the need to know the original position and the

accumulation of errors requiring continuous resetting of position using other sensors.

Landmark-based localization determines the absolute position of the device through

the recognition of predetermined distinct natural or artificial features of the

environment. Artificial landmarks are location reference markers attached to walls,

ceiling, or floor that can be easily recognized by vehicle-mounted instrumentation and

can be relatively inexpensive to install. Examples include special visual patterns [9],

infrared light-emitting diodes [10], and RFID tags [11-13].

Although both dead reckoning and landmark-based methods are error-prone,

fusion of the two can result in relatively reliable localization. Several localization

methods have been proposed that deal with uncertainty of the measurement data and

provide data fusion from different noisy sources such as dead reckoning and

landmarks. The Kalman Filter is a widely used method to compensate for noise and

is applicable to the localization problem [14]. Monte-Carlo or Particle Filter

localization [15] and fuzzy logic [16] have also been considered. In this paper we

present an algorithm for the fusion of RFID-based localization and inertial

measurement to obtain an accurate location of a delivery vehicle (forklift). With the

forklift already equipped with a mobile RFID reader, it is logical to consider the use

of passive RFID labels to create static location references in the environment. Since

the application considered does not require instantaneous knowledge of an asset's

location, the data from both inertial measurements and RFID are fused using post-

processing.

This paper describes the automotive part stamping environment and the need for

container tracking. Passive RFID and inertial-based localization are then presented

with strengths and weaknesses of each individually. The synergistic fusion of the two

methods is shown to eliminate the weaknesses of each and is further improved by

post-processing. The data fusion algorithm is then described, followed by

conclusions.

121

2 Background

The automotive stamping plant is a first tier supplier providing major components for

the vehicle body including doors, fenders, roofs, etc. The stamping plant supplies

parts to assembly plants and service facilities using truck or rail transportation. In

general, the stamping process consists of blanking, die press, and assembly/welding

operations. At the end of the press or assembly operations the parts are placed into

metal racks that are used to store and transfer them between stamping operations and

customers. Containers hold 8 to several hundred parts each and are unique to each

part. Typically, the size of the racks range between 4 to 12 feet in length, 3 to 7 feet

in width, 4 to 8 feet in height, and weigh up to 5000 lb when loaded. There are over

20,000 racks in any given plant, and for each given part type there is a limited rack

fleet. It is important to closely monitor the flow of racks within the plant and between

the plant and the customer's site. If the empty racks are not received back from the

customer on time or are held at the repair area, there may be an insufficient number of

racks to support production, resulting in non-optimal production batch sizes or

hampering the ability of the plant to satisfy customer demand. Each rack is tagged

with a passive RFID label, and fixed readers at shipping doors monitor rack flow

between the plant and its customers. However, more granular tracking using fixed

readers would require a substantial investment in infrastructure as a typical stamping

facility is over 1,000,000 square feet.

Fig. 1. Schematic of the flow of racks associated with the given work center.

The racks are handled by forklifts which move between the end of manufacturing

lines and storage areas, to and from trucks and rail cars, or to and from a repair area.

Figure 1 illustrates the rack flow within a typical stamping facility and associated with

a given set of parts produced by an assembly-welding operation (work center). The

Assembly

Work Center

Repair

Area

Customer A

Shipping Storage

Customer B

122

forklift transfers the racks with parts from the work center to the dedicated shipping

storage area and brings empty racks back to the work center. These parts can then be

loaded onto trucks or rail cars. If there are quality concerns regarding the parts stored

in the shipping area, the forklift may move these parts into a repair area and then

back. A typical stamping facility employs 40 forklifts, and each forklift is equipped

with a wireless terminal that can exchange data with a back office computer.

The many benefits of knowing the specific location and status of racks (full,

empty, or in repair) include saving time for material handling personnel locating

inventory and reducing downtime caused by unavailability of racks. In addition,

when parts are quarantined due to quality concerns, location information can reduce

the number of racks that are pulled from inventory. Location and status information

can also facilitate a first-in-first-out (FIFO) inventory control system, improve

inventory turnover rate, reduce the potential for obsolete parts, and improve material

flow and space utilization.

3 Inertial Tracking

Inertial measurement of motion involves the use of linear accelerometers and/or

rotational rate sensors whose signals are mathematically integrated to produce speed

and position estimates. These sensors are deployed in a wide array of applications

including air, space, and ground vehicles as well as various consumer electronic

systems. In each application, the orientation or dynamic motion state of the vehicle or

device is of interest. Automotive applications include anti-lock braking, traction

control, yaw and roll stability controls. Typical sensor sets for ground vehicles

measure longitudinal and lateral vehicle accelerations and yaw rate, with roll rate

seeing wider use now in roll mitigation systems. In automotive applications, vehicle

wheel speed sensors and GPS may also be added.

A general-purpose method for tracking a system in two-dimensional space

involves combining longitudinal and lateral accelerations and yaw rate. Yaw rate

(see Figure 2 for signal and axis definitions) is integrated to provide yaw angle or

heading:

∫

dtt

)(

ψψψ

. (1)

Vehicle-fixed longitudinal (x-axis) and lateral (y-axis) accelerations, a

and a

respectively, are then combined with heading and yaw rate to give velocities in the

fixed frame of reference (with respect to inertial ground):

[

]

∫

−−+=

yyxxx

dtVaaVtV

)sin()cos()(

ψψψ

, (2)

() ()

[

]

∫

+++=

xxyyy

dtVaaVtV

sincos)(

ψψψ

. (3)

123

Finally, these inertial velocities are integrated to give positions in the inertial frame:

∫

dtVXtX

)( , (4)

∫

dtVYtY

)( . (5)

Fig. 2. IMU signals and axes.

Because the sensor measurements are prone to offset and noise errors, this

double-integrating process results in error accumulation, limiting the time and

distance over which this process provides acceptable results. These limitations can be

mitigated to some degree by including kinematic constraints on the integration

process which are imposed by the vehicle itself. For instance, the forward speed, yaw

rate, and lateral acceleration of a wheeled vehicle are coupled directly as long as the

vehicle moves without sliding its wheels (a good assumption for heavy, factory floor

forklifts). Additionally, vehicle states such as speeds and rotational rates are bounded

by the vehicle operating envelope. The inertial tracking method benefits from a

continuous stream of data from the sensors, resulting in uninterrupted, fine-grained

position information. However, to anchor the path calculation in absolute space, the

position, heading, and velocity of the vehicle at the initiation of tracking (X

, Y

, V

x,0

y,0

, and

) must be known by some other means. For this reason, in addition to the

drift problem, inertial tracking alone is not completely suitable for forklift tracking.

4 RFID Grid Tracking

Unlike inertial tracking that drifts and has no inherent connection to absolute space,

RFID tags used as a tracking system can be physically connected to the operating

environment itself. RFID as a static positioning system requires the use of both tags

as well as a reading device that receives their identification information. The tags are

either attached at various fixed and known locations in the plant environment and

detected with a reader attached to the forklift, or they are deployed on the forklift and

124

read with a plurality of stationary readers. The former approach is advantageous as it

also facilitates the detection of the tagged and tracked inventory (racks). If the

location tags are deployed in a fine enough mesh within the plant environment,

forklift tracking can be accomplished using this method alone. However, despite their

low cost, passive RFID tags must still be installed, cataloged, and maintained and are

subject to damage in the hostile plant environment where suitable safe installation

locations may be few and far between. For example, a typical stamping plant may

have support columns, the only suitable installation location for tags, separated by 40

feet or more.

There are other inaccuracies as well. Because the tags can be read at a distance

from a range of orientations, the exact location of the reading device attached to the

tracked vehicle is not known when communication is acquired with a tag. Therefore,

the static RFID tag location signal has some position error since the passing reading

device can only be assumed to be located within some expected range (10 to 15 feet)

of the energized tag. Additionally, there may be time when the tracked vehicle is not

in range of any tag or where a tag is damaged and inoperative. Thus the RFID

tracking method, while attached to inertial space, is nevertheless somewhat

inaccurate, discrete in nature, and intermittent.

5 IMU and RFID Fusion

Combining the inertial (IMU) and static (RFID) tag data can reduce or eliminate the

limitations associated with each individually. The inertial measurements provide

continuous data that is potentially accurate enough to reasonably "connect the dots"

between the sparse static tag data. Using the absolute position data from the RFID

tags, offset errors in the IMU readings can be estimated and removed, resulting in

improved positioning, and the IMU data integration can be initialized with static

RFID location information to connect it to absolute space.

In the automotive stamping plant setting, the instantaneous location of a tracked

vehicle is less important than an accurate estimate of the vehicle's path during a

finished delivery as each event takes a few minutes or less. Therefore, the method

proposed here assumes that inertial sensor data and static tag data will be collected

and stored during an event. When the event ends, the data will be post-processed to

determine the path of the delivery vehicle (especially its end points). Using data from

the entire event provides a richer data set from which to calculate path, whereas

attempting a continuous, immediate position estimate during the delivery event limits

the calculations to data that has occurred in the past only and provides no real benefit

as the post-processed data is timely enough (within a few seconds of the event

ending) for the stamping environment.

6 Postprocessing using Best-Fit Optimization

The proposed computational method relies upon a recursive solution that makes

repeated guesses of the initial vehicle position and IMU sensor offsets to attempt to

125

minimize the error between the resulting IMU-based path estimate and the known

locations of the static RFID position readings taken over that path. As the solution

converges, RFID reader range information is employed to attempt to predict the RFID

signal acquisition and loss locations to further improve solution fidelity. The full

method is consists of the following 6 steps.

Step 1. Inertial (IMU) and static tag (RFID) data is collected during an inventory

delivery event. Inertial data is collected continuously at a regular rate, typically 10 to

100 Hz. Static tag data consists of the time and ID number of each tag as

communication with the tag is acquired and then lost.

Step 2. Upon completion of the delivery event (determined by observing acquisition

and loss of inventory tag readings), the vehicle path is reconstructed solely from the

inertial data by making vehicle initial state assumptions (X

, Y

, V

x,0

, V

y,0

, and

Typically, these initial conditions are chosen based on information from the end of the

previous delivery event.

Step 3. Using a best guess position for each RFID tag reading, one that represents the

most likely average vehicle position while it is in communication range with the

location tag, a path error calculation is made using the path constructed in Step 2:

[][]

∑

−+−=

nRFIDnnRFIDn

YtYXtX

)()(

. (6)

Here n is an index indicating the acquisition and loss events for the RFID tag

readings, and X(t

) and Y(t

) give the position of the IMU integration calculation

(Equations 1-5) at these RFID tag detection times. The best guess positions, X

RFID,n

and Y

RFID,n

, are unique for each tag and are based on the position of the tag, the most

likely vehicle path followed when in range of the tag, and the reader's expected

communication range. These values can be cataloged during the tag's installation or

constructed from data collected during system operation.

Step 4. The assumed vehicle initial state (first used in Step 2) is perturbed, and Steps

2 and 3 are repeated until these parameters converge to minimize the error calculated

in Step 3. This iterative process can be conducted using any of a number of

optimization algorithms such as Matlab's fminsearch function. The result of this step

is the best match of the inertial sensor-based position to the best guess RFID tag

positions used in Step 3.

Step 5. The result of Step 4 is used to calculate the most likely RFID tag acquisition

and loss locations using detection range assumptions. At each time of acquisition or

loss, the vehicle position (from Step 4) and tag location data are used to find the

intersection of the estimated vehicle path with the locus of expected tag detection

range points. This locus can be assumed to be circular, or more complex shapes can

be used based on more detailed tag and reader information. These new

126

acquisition/loss points establish a more likely vehicle path for further optimization of

the vehicle initial conditions (X

, Y

, V

x,0

, V

y,0

, and

Step 6. Steps 3 and 4 are repeated using the new acquisition/loss positions produced

in Step 5. These positions are updated with each iteration of the inertial data path

optimization, and this step is repeated until the solution converges and each

subsequent iteration produces a path prediction that is negligibly different from the

iteration before.

Fig. 3. Forklift path reconstruction showing actual forklift position evenly spaced in time,

RFID antenna locations and ranges, initial path estimate, and final path estimate after ten

iterations.

Figure 3 illustrates several of the aspects of the method described above. The

actual path of the forklift is shown for a delivery event with forklift orientation

superimposed at regular time intervals. Large solid circles surround the RFID

location tags and depict their actual detection range, while the large dashed circles are

the assumed reader ranges used in Step 5. The smaller circles located within the tag

-20 -10 0 10 20 30

-40

-35

-30

-25

-20

-15

-10

-5

X-Position, [m]

RFID senso

locations

Actual vehicle path

Path estimate based on

initial best guess sensor

locations (Step 4)

Initial best guess

sensor locations

(Step 3)

Actual static

sensor acquisition

& loss

oints

Optimized vehicle path

(Step 6) after 10 iterations

Static sensor assumed

(dotted line) and actual

(solid line) communication

ranges

Calculated sensor

acquisition/loss points

(Step 5) after 10 iterations

Y-Position, [m]

127

detection range are the initial best guess detection locations used in Step 3, and the

half-tone line is the path resulting from the optimization using these points (result of

Step 4). Shown also is the result of the 10th iteration of Step 6 using updated

acquisition and loss points calculated in Step 5. Note the convergence of the

optimized path toward the actual path from the Step 4 result to the Step 6 result.

The accuracy of the path reconstruction depends mostly on the accuracy of the

assumed reader detection range as compared to the actual range. Note from the figure

that the path error (the distance between the actual and final estimated path) is

roughly the same size as the reader range variability. For a typical forklift-mounted

RFID reader, this is 5 – 10 feet. Solution accuracy can be further improved by

including inertial sensor offset errors that are optimized with the initial conditions.

7 Conclusions

Presented here is an effective method of indoor localization of forklift vehicles

equipped with mobile RFID readers. The method is based on the fusion of inertial

measurements with information from static RFID location tags. While each individual

method alone is prone to substantial errors, fusion of the two can provide results of

reasonable quality while minimizing the cost of implementation. The accuracy of the

resulting location estimate depends on the spacing of the location tags throughout the

plant as well as the variability of the RFID reader detection range. In practice it is

possible to achieve an accuracy within 5-7 feet which is acceptable for the stamping

plant problem.

The ultimate goal of this development is to determine and record the locations of

the RFID tagged containers handled by a given forklift. Specifically, the described

approach has been developed to track the rack locations for automotive stamping

plants. While this paper has focused on the fusion of the inertial measurements with

the static information from the RFID tag locations for the purpose of tracking the

delivery vehicle, it should be noted that the full material tracking problem is more

complex. In addition to location awareness, the pick-up and delivery event with its

associated inventory must be identified by monitoring the stream of RFID tag

readings seen by the mobile reader. This must be accomplished robustly despite

delivery complexities such as moving stacks of racks where not all of the moved

racks will be seen by the reader during the entirety of the delivery event.

References

1. Thiesse, F., Fleisch, E.: On the Value of Location Information to Lot Scheduling in

Complex Manufacturing Processes, Int. Journal of Production Economics, Vol. 112, 2008,

pp. 532-547.

2. Jones, E.C., Chung, C.A.: RFID in Logistics: A Practical Introduction, CRC Press, 2008.

3. Sanpechuda, T., Kovavisaruch, L.: A review of RFID localization: Applications and

Techniques, Proceedings of 5th International Conference on Electrical

Engineering/Electronics, Computer, Telecommunications and Information Technology

ECTI-CON 2008, pp. 769 – 772.

128

4. Jungk, A., Heiserich, G., Overmeyer, L.: Forklift Trucks as Mobile Radio Frequency

Identification Antenna Gates in Material Flow, Proceeding of Conference on Intelligent

Transportation Systems, 2007, pp. 940 – 943.

5. Michel, J. C. F., Millner, H., Vossiek, M.: A Novell Wireless Forklift Positioning System

for Indoor and Outdoor Use, 5th Workshop on Positioning, Navigation and Communication

2008 (WPNC’08), Hannover, Germany, March 2008, pp. 219-227.

6. Röhrig, C., Spieker, S.: Tracking of Transport Vehicles for Warehouse Management using

a Wireless Sensor Network, Proceedings of IEEE/RSJ International Conference on

Intelligent Robots and Systems, Nice, France, September 22-26, 2008, pp. 3260-3265.

7. Borenstein, J., Everett, B., Feng, L.: Navigating Mobile Robots: Systems and Techniques,

A. K. Peters, Ltd., Wellesley, MA, 1996.

8. Hu H., Gu D.: Landmark-based Navigation of Industrial Mobile Robots, International

Journal of Industry Robot, Vol. 27, No. 6, 2000, pp. 458 – 467.

9. Yoon, K.-J., Kweon, I. S., Lee, C.-H., Oh, J.-K., Yeo, I.-T.: Landmark Design and Real-

Time Landmark Tracking Using Color Histogram for Mobile Robot Localization,

International Symposium on Robotics (ISR), 2001, pp. 1083-1088.

10. Lee, S., Song, J.-B.: Use of coded infrared light as artificial landmarks for mobile robot

localization, Proceeding of International Conference on Intelligent Robots and Systems,

2007, pp. 1731-1736.

11. Hahnel, D., Burgard, W., Fox, D., Fishkin, K., Philipose, M.: Mapping and Localization

with RFID Technology, Proceedings of ICRA 2004, pp. 1015-1020.

12. Schneegans, S., Vorst, P., Zell, A.: Using RFID Snapshots for Mobile Robot Self-

Localization, Proceedings of the 3rd European Conference on Mobile Robots (ECMR

2007), Freiburg, Germany, 2007, pp. 241-246.

13. Yang, P., Wenyan, W., Moniri, M., Chibelushi, C.C.: SLAM Algorithm for 2D Object

Trajectory Tracking based on RFID Passive Tags, Proceeding of 2008 IEEE International

Conference on RFID, 2008, pp.165 – 172.

14. Santana, A.M., Sousa, A. S., Britto, R.S., Alsina, P.J., Medeiros, A.D.: Localization of a

Mobile Robot Based in Odometry and Natural Landmarks Using Extended Kalman Filter,

Proceedings of 5

International Conference on Informatics in Control, Automation and

Robotics, 2008, pp. 187-193.

15. Thrun, S., Fox, D., Burgard, W., Dellaert, F.: Robust Monte Carlo Localization for Mobile

Robots, Artificial Intelligence, Vol. 128, 2001, pp. 99-141.

16. Demirli, K., Molhim, M.: Fuzzy dynamic localization for mobile robots, Fuzzy Sets and

Systems, Vol. 144, 2004, pp. 251–283.

129