08-16-2017, 09:44 PM
SixthSense: RFID-based Enterprise Intelligence
ABSTRACT
RFID is widely used to track the movement of goods through
a supply chain. In this paper, we extend the domain of
RFID by presenting SixthSense, a platform for RFID-based
enterprise intelligence systems. We consider an enterprise
setting where people (or rather their employee badges) and
their personal objects such as books and mobiles are tagged
with cheap, passive RFID tags, and there is good coverage of
RFID readers in the workplace. SixthSense combines mobil-
ity information obtained from RFID-based sensing with in-
formation from enterprise systems such as calendar and pres-
ence, to automatically draw inferences about the association
and interaction amongst people, objects, and workspaces.
For instance, SixthSense is able to automatically distinguish
between people and objects, learn the identities of people,
and infer the ownership of objects by people.
We characterize the performance of a state-of-the-art RFID
system used in our testbed, present our inference algorithms,
and evaluate these both in a small testbed and via simula-
tions. We also present the SixthSense programming model
that exposes a rich API to applications. To demonstrate
the capabilities of the SixthSense platform, we present a
few applications built using these APIs, including a mis-
placed object alert service, an enhanced calendar service,
and rich annotation of video with physical events. We also
discuss the issue of safeguarding user privacy in the context
of SixthSense.
Categories and Subject Descriptors:
C.m [Miscellaneous]: Sensing Systems
General Terms:
Design, Experimentation, Measurement.
Keywords:
RFID, sensing, ubiquitous computing.
The author was an intern at Microsoft Research India dur-
ing the course of this work.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
MobiSys 08, June 17 20, 2008, Breckenridge, Colorado, USA.
Copyright 2008 ACM 978-1-60558-139-2/08/06 ..$5.00.
1. INTRODUCTION
Radio Frequency Identification (RFID) [15, 27, 13] is an
electronic tagging technology that allows the detection and
tracking of tags, and consequently the objects they are af-
fixed to. An RFID tag typically comprises a passive transpon-
der that responds with identifying information when ener-
gized remotely by an RFID reader. This ability to do re-
mote detection and tracking coupled with the low cost of
passive tags has led to the widespread adoption of RFID in
the supply chain world. RFID is used to track the move-
ment of goods through a supply chain, whether it be pallets
shipped between warehouses, cases delivered to stores, or
items placed on the store shelves, thereby optimizing inven-
tory management and yielding significant cost savings.
The promise of cheap connectivity to any object carry-
ing an RFID tag has led to the vision of an Internet of
Things [11, 26]. Our work on SixthSense is inspired by
this vision. SixthSense focuses on applying RFID to an
enterprise setting, such as a corporate office or university
department. An enterprise setting is different from the sup-
ply chain scenario in one fundamental way: the central role
of people. Unlike in a supply chain, an enterprise setting
involves rich interaction amongst people, between people
and objects, and between people and workspaces. For in-
stance, people own objects such as books, cell phones, and
laptops, which they often carry around and sometimes mis-
place. SixthSense provides a platform for tracking and infer-
ring such interactions, and then exposing these to the higher
layers via APIs that enable useful applications and services
to be built. Thus SixthSense raises the level of abstraction
for applications in this domain beyond tag-level events, akin
to how RFID stacks such as Microsoft s BizTalk [7] do so in
the supply chain context. In short, SixthSense represents a
form of mobile computing applied to non-computing entities.
SixthSense assumes a setting where most people (or rather
their employee badges) and objects are tagged with passive
RFID tags, and the coverage of RFID readers spans much
of the workspace. However, we do not assume that this
tagging is always catalogued systematically. Indeed, many
objects present in a workplace may not even belong to the
enterprise (e.g., a user s personal mobile phone). Even if all
objects (and people) were cataloged, this would be a man-
ual process prone to errors and furthermore would require
updating each time a new object is added or an object needs
to be retagged because of the deterioration of its old tag [8].
Therefore, a key goal of SixthSense is to make all inferences
automatically, without requiring any human input. Even in
settings where human input is available, the inference algo-
rithms in SixthSense can help catch errors, e.g., the wrong
ownership information for an object being recorded in a cat-
alog.
SixthSense incorporates algorithms that start with a mass
of undifferentiated tags and automatically infer a range of in-
formation based on an accumulation of observations. Sixth-
Sense is able to automatically differentiate between people
tags and object tags, learn the identities of people, infer the
ownership of objects by people, learn the nature different
zones in a workspace (e.g., private office versus conference
room), and perform other such inferences. Mobility of peo-
ple and objects is key to the inference performed by Sixth-
Sense. For example, tags attached to people are more likely
to move, with less dependence on other tags, than tags at-
tached to objects. Likewise, the owner of an object is likely
to be the person who carries it around the most.
Since RFID by itself only provides very limited informa-
tion basically, just the presence or absence of a tag in
a particular zone SixthSense also leverages information
from other enterprise systems, e.g., calendar, presence, lo-
gin information, etc. By combining information from these
diverse sources, SixthSense records all tag-level events in
a raw database. The inference algorithms consume these
raw events to infer events at the level of people, objects,
and workspace zones, which are then recorded in a sepa-
rate processed database. Applications can either poll these
databases (e.g., by running SQL queries) or set up triggers
to be notified of specific events of interest. We present a
few applications that we have implemented on top of Sixth-
Sense: lost object alert, enhanced calendar and presence,
semi-automated image cataloging of objects, and rich anno-
tation of video with physical events.
We envision SixthSense being run centrally by the en-
terprise rather than by individual users. This includes the
raw and processed databases, and the applications that con-
sume the information contained in these databases. While
such a model limits flexibility, it greatly simplifies deploy-
ment issues, specifically with regard to privacy. Individual
user s are unable to access the SixthSense databases; they
are only presented with information that the centrally-run
application chooses to expose, e.g., alerts regarding a user s
own objects that have been misplaced. We also employ a
simple tag relabeling scheme to defeat any attempts to re-
construct the database surreptitiously, say using input from
rogue readers. While the enterprise itself will have access to
potentially privacy-sensitive data, this is not fundamentally
different from the present situation, where the enterprise
has access (with legal sanction, in some countries [17]) to
arguably more sensitive information such as the employees
email and files (largely in an unencrypted form), and indeed
also the ability to track user movements to an extent based
on card key based access control (which is, in fact, based
on short-range RFID) to various physical spaces. We be-
lieve that the safeguards in place to guard against leakage
or abuse of such sensitive information could be extended to
SixthSense.
In summary, the main contribution of our work is the de-
sign, implementation, and evaluation of SixthSense, a plat-
form for RFID-based enterprise intelligence that combines
RFID events with information from other enterprise systems
and sensors to automatically make inferences about people,
objects, workspaces, and their interaction.
Class 0 Passive Read only
Class 1 Passive Read only write once
but with rewritable 96-bit EPC
Class 2 Passive 65 KB read-write
Class 3 Semi-passive 65 KB read-write
with built-in battery
Class 4 Active Built-in battery
Class 5 Active Communicates with other
class 5 tags and devices
Table 1: Different classes of tags.
2. RFID BACKGROUND
Radio Frequency Identification (RFID) [15, 27, 13] has
been around for decades. It is generally believed that the
roots of RFID can be traced back toWorldWar II [10], when
the British first put transmitters on their aircraft, which on
receiving radar signals, broadcast back a signal to ground
station identifying the plane as a friend. Since then, the
technology has improved tremendously and RFID has seen
large deployments, especially in the supply chain and asset
tracking domains.
An RFID system comprises a reader, with one or more
antennas attached to it, and tags. When a reader energizes
an antenna, the tags in the corresponding zone get activated
and respond with their ID (e.g., a 96-bit electronic product
code (EPC)) and possibly other data.
RFID tags come in three types: passive, active and semi-
passive. Passive tags do not have any internal power supply.
Instead, they use the electric current induced in the tag s
antenna by the incoming RF signal from a reader to power
the tag s IC and transmit a response. Such tags typically
have a read range of about 10 cm up to a few meters. The
simplicity of these tags has also meant a low cost about
15 U.S. cents per tag today, expected to go down to 5 U.S.
cents [9]. Active tags, on the other hand, have their own
internal battery to power the IC and transmit a response
using an arbitrary RF technology such asWiFi. They have a
read range of hundreds of meters, due to the internal battery.
Semi-passive tags have their own power supply to power the
IC and to help with reception, but like passive tags they use
the RF induced current for transmitting a response back to
the reader.
Passive tags typically receive power through inductive or
radiative coupling. Inductive coupling is used for powering
LF (low frequency, 30-300kHz) and HF (high frequency, 2-
20 MHz) tags. Such tags receive power in the near field,
which refers to the region within a few wavelengths of the
reader s antenna. A reader antenna generates a magnetic
field, inducing an electric current in the tag s antenna and
charging a capacitor in the tag. Radiative coupling is used
for UHF tags (Ultra High Frequency, above 100 MHz). In
this case, the tag antenna receives signals and energy from
the electromagnetic field emitted by the reader in the far
field, the area beyond a few wavelengths.
With increasing RFID deployments, a need for standard-
ization was felt for ensuring interoperability of the RFID sys-
tems from different vendors. The Auto-ID Center at MIT,
which is now being managed by EPCglobal [1], developed
the electronic product code (EPC). EPCglobal has defined
different classes of tags, as shown in Table 2.
Class-0 and Class-1 tags are not interoperable and they
are not compatible with ISO standards. In 2004, EPCglobal
began developing a Class-1 Generation-2 protocol (Class-1
Gen-2 or just Gen-2), which would not be backward com-
patible with either Class-1 Generation-1 (Class-1 Gen-1) or
Class-0 tags. The aim was to create a single, global stan-
dard that would be more closely aligned with ISO standards.
Class-1 Gen-2 was approved in December 2004.
With the increasing use of RFID technology for retail sys-
tems, there has been a concern that the privacy of individu-
als that purchase items would be jeopardized by the ability
to identify items uniquely and surreptitiously. Responding
to this concern, the Auto-ID center, and later EPCglobal,
included an option to kill a tag after purchase if a customer
desires to protect their privacy. For example, in Class-0 tags,
a 24-bit password is programmed into the tag during man-
ufacturing. The password can be used to kill the tag. Once
a reader accesses a tag with the correct 24-bit password and
issues the kill command, a fusible page link on the tag becomes
open, rendering it unreadable.
3. EXPERIMENTAL SETUP
We briefly discuss our experimental setup.
3.1 RFID Equipment
We used an Impinj Speedway reader [4] for our experi-
ments. This is a state-of-the-art UHF Class 1 Gen 2 reader
that is compliant with EPCglobal and ISO standards. It
operates in the 865-956 MHz band. Our setup is equipped
with 4 patch antennas, each measuring 26 cm by 26 cm. The
RF power output by the antennas is +30 dBm (1 W). The
receiver sensitivity of the reader is -80 dBm.
We tag objects with Impinj Monza passive RFID tags [3].
These are UHF Class 1 Gen 2 tags that have been certified
by EPCglobal. It includes a 96-bit field-rewritable EPC and
supports a 32-bit password-protected kill command. The
tag, including the chip and packaging, measures 9 cm by 5
cm.
The reader connects to the corporate network using Eth-
ernet and the host computer connects to the reader using
its IP Address. The reader can be connected to 4 RFID
antennas. The Speedway reader exposes the RSSI of a tag
being read in addition to its EPC. The reader also has the
capability to write tags.
Figure 1 shows a picture of the reader with its antennas
and a few objects that were tagged.
3.2 Physical Setup and Enterprise Setting
Our work is set in the Microsoft Research India Lab.
We deployed our RFID reader to cover a section of the
workspace on one floor measuring 10 meters by 6 meters.
This space is occupied by 4 users, who served as the test sub-
jects for our experiments. We tagged several objects belong-
ing to each user: their employee badges, mobile phones, lap-
tops, books, water bottles, etc. These users used a calendar
system hosted on Microsoft s corporate Exchange servers.
Also, they logged in to a corporate domain and signed in to
a Microsoft Universal Communicator presence service, both
of which indicated user activity at their computer. 1
Figure 2 shows a view of the workspace, including the
1We would want to filter out login or sign-in activity per-
formed remotely, which would be straightforward to do.
Figure 1: Speedway Reader, Antenna and Monza
Tags
Figure 2: RFID deployment
RFID reader antennas and the users along with their tagged
objects. While the antennas are placed on the users desk in
the current setup, our eventual plan is to have these mounted
on the ceiling.
3.3 SixthSense Simulator and Visualizer
Given the small size of our testbed, we developed a sim-
ulator to enable experiments at larger scale. The simulator
incorporates simple models of object ownership, user mobil-
ity (possibly carrying one or more objects), objects being
misplaced, user logins, etc. The simulator then generates
a synthetic trace of RFID and other events, which is then
fed into the SixthSense system for analysis and inference.
Our tool also incorporates a visualizer that depicts users,
objects, and their movements.
4. RFID PERFORMANCEMEASUREMENTS
While there have been many measurement studies of RFID,
as discussed in Section 11, we would like to characterize the
performance of the Impinj Speedway reader in our particu-
lar setting. To this end, we present some basic performance
measurements.
Figure 3: Measurement setup
700
800
600
Maximum Read Range (cm)
500
400
Parallel
200
300
Perpendicular
100
0
0 50 100 150 200 250
Displacement from Antenna Axis
Figure 4: Effect of displacement from antenna axis
on read range
300
250
Maximum displacement from axis (cm)
200
150
Parallel
100
Perpendicular
50
0
0 80 180 260
Distance from Antenna (cm)
Figure 5: Effect of distance from antenna on allowed
displacement from axis
Figure 6: Read ranges for tags on objects carried by
a person
In the first set of experiments, we studied the attainable
read range for different displacements of the tag from the
normal axis of the antenna. Figure 3 shows the measure-
ment setup. At each displacement from the antenna axis,
the maximum read range is recorded as the farthest distance
at which the tag can still be read. It can be inferred from
Figure 4 that as the tag moves away from the antenna axis,
the read range decreases. Also, when a tag is in a perpendic-
ular orientation with respect to the plane of the antenna, the
read range is lower. In another experiment, we kept the tag
at different distances from antenna and measured the maxi-
mum allowable displacement from the antenna axis that still
allowed the tag to be read. Figure 5 shows that as the tag
moves farther from antenna, the allowed displacement from
axis also increases proportionally. Also, a tag in perpendicu-
lar orientation allows a much smaller displacement from axis
as compared to a tag in parallel orientation.
We also noted that when a tag is in the range of the reader,
the read rate or response rate (i.e., percentage of times a
tag is read when probed) is 100% and when it is outside the
range, the read rate immediately falls to 0%. This behav-
ior is in sharp contrast to the more gradual degradation in
read rate at the read range boundary, as reported in other
studies [22, 23]
We then characterized the read ranges for tags affixed to
objects such as cell phone, wallet, books and laptops, which
are often carried by people in an enterprise environment.
Figure 6 shows the observed read ranges for different ob-
jects in both parallel and perpendicular orientations. Tags
affixed to books have a read range similar to that of tags in
the open, i.e., not affixed to any object. An interesting ob-
servation was the effect of insulation layers on tags attached
to metallic bodies (e.g., laptops). We observed that without
any insulation, such tags had very low read ranges (50 cm).
However, with just 3 layers of thin card board insulation, we
were able to obtain read ranges of 150 cm. In general, our
experiments showed that the read range increases with an
increasing number of insulation layers. It can also be noted
from Figure 6 that the read range is lower when the tag is
oriented perpendicular to the antenna as compared to when
it is oriented parallel to the antenna.
Setting the transmission power of the reader to different
values, we experimented with the read ranges obtained. We
observed that as the power increases, the read range also in-
creases. With the highest power level, the observed antenna
range was about 750 cm (7.5 m), which is large enough to
cover most rooms in an office building. This is the power
setting we used in all of the SixthSense experiments.
[b]5. SIXTHSENSE ARCHITECTURE[/b]
We now lay out the overall architectural rationale and
structure of SixthSense.
5.1 Assumptions
SixthSense assumes an enterprise setting with widespread
coverage of RFID readers, and where most or all people
and objects are tagged with passive RFID tags. However,
we do not assume that the tagging of people and objects
is cataloged. Users are free to pick up new tags and affix
them to objects, as and when needed. This low-overhead
model with little control is appropriate for SixthSense be-
cause these long-range UHF tags do not serve any security
function, unlike the short-range HF tags embedded in em-
ployee card keys.
We also assume that users have access to a computing
environment that provides services such as network logins,
shared calendars, and online presence, which can be mon-
itored by SixthSense. This is increasingly the case in en-
terprises, with the adoption of networked systems such as
Microsoft Exchange and IBM Lotus Notes.
5.2 Architectural Components
Figure 7 shows the SixthSense architecture. The key com-
ponents of the system, including the databases, inference
engine, and applications, are run centrally by the enter-
prise. This provides the (trusted) inference engine access
to the complete set of sensed data across all users, objects,
and zones, allowing it to make effective inferences. Like-
wise, the (trusted) application is allowed the flexibility of
working with a complete set of inferences (i.e., inferences
pertaining to all users and their objects), yet control what
processed information is presented to the users to ensure pri-
vacy. In contrast, if the inference engine or the application
were run by individual users on their own desktop machines,
privacy consideration would restrict the set of information
made available to these, and hence limit their functional-
ity. For example, privacy considerations would disallow an
application run by one user from accessing inferences per-
taining to another user, making it difficult to implement new
functionality such as the automatic conference room book-
ing feature discussed in Section 9.3.
Next, we briefly discuss the various components of Sixth-
Sense.
5.2.1 RFID Monitor
The RFIDMonitor issues a read command every 500 ms to
the RFID reader. The reader reports the EPC and the signal
strength (RSSI) of the tags read via each of its antennas.
This data then gets pushed into the raw database.
5.2.2 Other Enterprise Monitors
These other monitors monitor the information listed below
and push their updates into the the raw database:
Calendar Monitor: This resides on each user s desktop
machine, and monitors the time and location of the
user s appointments.
Figure 7: SixthSense Architecture
Presence Monitor: This monitors the status of each
user s interaction with their desktop. A machine is said
to be idle when it receives no user input for 2 minutes.
Transitions from idle to active state are detected and
reported.
Login Monitor: This is similar to the presence monitor
except that in general login is a much stronger indica-
tion of a user being present than simply a change in
their machine s idle.
Cameras: Office buildings are often equipped with
cameras for security reasons. The camera feed is stored
in a video database for future analysis, if the need
arises. We show in 9.2 how we combine the camera
system in an enterprise with other sensors to build
useful applications.
5.2.3 Raw database
The RFID monitor and the other enterprise monitors push
data into the raw database.
5.2.4 Inference engine
The inference engine operates on the raw database us-
ing the algorithms explained in Section 6 to draw inferences
about people, objects, and workspaces.
5.2.5 Processed database
The processed database is populated by the inference en-
gine with its inferences, making these available to applica-
tions built on top of the SixthSense platform.
5.2.6 API
SixthSense provides a set of APIs for applications to lookup
the inferences stored in the processed database or to receive
callbacks when new inferences are made. We elaborate on
this in our discussion of the SixthSense programming model
in Section 8.
5.2.7 Applications
A range of applications can be built using the APIs ex-
posed by the SixthSense platform. We discuss a few that we
have built in Section 9.