1

Cardinality Estimation for Large-scale RFID Systems

Chen Qian, Hoilun Ngan, and Yunhao LiuHong Kong University of Science and Technology

2

RFID: Hot Topic

• Both in industry and academic society

• RFID: independent sessions (three or more papers) in PerCom 2007, 2008

• 2009?

3

Research Issues(take our group as an example)

• Localization

• Object Tracking

• Security & Privacy

• Tag Counting & Estimation

To be expanded…

4

RFID: Hot Topic

• Some RFID papers in other top confs.M. S. Kodialam, T. Nandagopal, “Fast and reliable

estimation schemes in RFID systems”, MobiCom 2006J. Myung, W. Lee, “Adaptive splitting protocols for RFID tag collision arbitration”, MobiHoc 2006

Qunfeng Dong, et. al., “Load Balancing in Large-Scale RFID Systems”, Infocom 2007

Z. Zhou, et. al., "Slotted Scheduled Tag Access in Multi-Reader RFID Systems", ICNP 2007

5

RFID Sys. Model

RFID Readers– Carrying

antennas, collect info from nearby tags.

– Connected with servers

RFID Tags– Labeled with

unique serial #s– Simple structure– Large-deployed,

but can not communicate with each other

If multiple tags transmit to reader simultaneously, a collision happens, and reader cannot recognize these tags.

6

Real Problems

• RFID tags are used to label large-volume items.

• Hence, collecting the information of these items is the main goal of the RFID system.

• Two main kinds of information: – Identities Cardinality

Identification Counting

9

Tag counting:Some applications

• Hong Kong International Airport

Cargo transportations

10

Tag counting:Some applications

• Stadium RFID System

Security and traffic control

11

Identification:Limitation

• We can obtain the tag cardinality via identification.

But….

Extremely long latency– 1000 sec for 3000 tags

Not applicable for mobile objects

12

Estimation(Mobicom 06)

13

Estimation:Limitation

• Multiple-reading problem

14

Our Goal

• Design an estimation scheme that can– Eliminate replications from the sum of

reader results. – Achieve a short processing time,– And high accuracy.

15

LPE

• Linear Probabilistic Estimation (LPE)

Replication-insensitive

16

LPE:Limitation

• Processing time is still too long to be ideal

• One can never know in advance that how long the ALOHA frame should be set.

17

Can we design an estimation schemethat works well without pre-knowledge?

18

Galton Board

19

GD Galton Board

Geometric Distribution

20

LoF

• Lottery Frame

Approximately1/2(t+1) of the tag responses are in time slot t.

21

LoF

The kth bit in bitmap BM[k] will be zero if k>>log2n, or be one if k<<log2n.

The fringe consists zeros and ones for the k whose value is near log2n.

2Rn R is the position of the right most zero

22

LoF

P. Flajolet and G. N. Martin, "Probabilistic Counting Algorithms for Data Base Applications," Journal of Computer and System Science, vol. 31, 1985.

1.2897

1.2897 2Rn

23

LoF:accuracy

• LoF estimation may not be accurate enough for some applications.

• Luckily the right most zero R is an unbiased estimator of log2n, which meansIf we make several independent estimations

and compute the average result, the standard error will be reduced.

24

LoF:multiple hashes

1 2 ... mR R RR

m

R

Consider the average value

The variable has the expectation and standard deviation that satisfy

( ) /R m

Therefore, the improved estimator is /

1.2897 2 1.2897 2i

i

R mRn

25

LoF:accuracy

26

LoF:processing time

The number of time slots required for a frame is independent from the size of tag set.

A frame with 16 slots is enough toestimate up to 216 = 65536 tags.

27

Simulation:setup

Fixed 32-slot length for LoF estimation.

28

Simulation:Single reader

29

Simulation:Single reader

0 10000 20000 30000 40000 500000.1

1

10

Number of Tags

No

rma

lize

d S

tan

da

rd D

evi

atio

n LoF (1 hash)LoF (2 hashes)LoF (4 hashes)LoF (8 hashes)LoF (16 hashes)LoF (32 hashes)

30

Simulation:Multiple reader

50 55 60 65 700

0.25

0.5

0.75

1

Reading Range (units)

Err

or

UPE (Max)UPE (Sum)LoF (16 hashes)LoF (64 hashes)

31

Simulation:Processing time

Just the last time!

32

Summary

• LoF is a replication-insensitive estimation, working well in multi-reader environments.

• LoF can obtain higher accuracy and lower latency, comparing with previous schemes.

• Trade-off in LoF: the storage for hash functions.

33

• Questions?

Thank you !