A Hybrid Method for achieving High A Hybrid Method for achieving High Accuracy and Efficiency in Object Tracking Accuracy and Efficiency in Object Tracking

using Passive RFIDusing Passive RFID

Lei Yang1, Jiannong Cao1, Weiping Zhu1, and Shaojie Tang2

1Hong Kong Polytechnic University, Hong Kong2Illinois Institute of Technology, USA

OutlineOutline

Introduction

Existing Approaches

Our Approach

Evalution

Conclusions

MotivationMotivation

Technical DifficultiesTechnical DifficultiesReal-time characterstic

Tracking is more difficult than localization of stationary objects because executiion the algorithm needs to be finished before a deadline

Noisy measurementRFID reading is noisy, which means the tags have low probability to be detected by the reader even though they are within the object’s reading range

Constrained computional resource on the mobile RFID devices

Problem DefinitionProblem DefinitionGiven

K tags deployed in the enrionment

The location of the i-th tag Ti

The RFID reader scans frequency f

The reader’s reading range is tr , which can be adjusted

The RFID reading at time k

AssumeThe reader has a probability p(r) to read the tags within its reading range

ObjectiveEstimate the continuous locations of the mobile object using the uncertain RFID readings

H

tr

Antenna

Tags

d

v

1{ }i Kk k iz r

, ( )

0,

re r trp r

r tr

Existing ApproachesExisting Approaches

Centroid Localization (CL) [N.Bulusu 2000]

Average the locations of all the tags that have been detected by the RFID readerThe tracking problem is solved by executing CL periodicallyLow accuracy in case of low detecting probability

Weighted Centroid Localization (WCL)Improve the accuracy by assigning weights in averaging the tags’ locationEach tag’s weight equals to the times that the tag has been detected in past N scansLarge error if the object’s speed is large relative to the scanning speed of the RFID reader

Both methods are computationally cheap

Estimated Location

Estimated Location

Existing ApproachesExisting Approaches

Centroid Localization (CL) [N.Bulusu 00]

Average the location of all the tags that have been detected by the RFID readerThe tracking problem is solved by executing CL periodicallyLow accuracy in case of low detecting probability

Weighted Centroid Localization (WCL) [Behnke08]

Improve the accuracy by assigning weights in averaging the tags’ locationEach tag’s weight equals to the times that the tag has been detected in past N scansLarge error if the object’s speed is large relative to the scanning speed of the RFID reader

Both methods are computationally cheap

• Count the detection times for each tag in recent 5 scans;• The weight of each tags is equal to the count number.

t = 1

t = 2

t = 3

t = 4

1 4 2

1 3 2

Estimated Location

t = 5Detection times

v

• WCL has large error if the object’s speed is large relative to the scanning speed of the RFID reader .

Existing ApproachesExisting Approaches

Centroid Localization (CL) [N.Bulusu 00]

Average the location of all the tags that have been detected by the RFID readerThe tracking problem is solved by executing CL periodicallyLow accuracy in case of low detecting probability

Weighted Centroid Localization (WCL) [Behnke08]

Improve the accuracy by assigning weights in averaging the tags’ locationEach tag’s weight equals to the times that the tag has been detected in past N scansLarge error if the object’s speed is large relative to the scanning speed of the RFID reader

Both methods are computationally cheap

Existing ApproachesExisting Approaches

Particle Filter [D.Hahnel06 ][Schneegans07] ][Vorst08]

The object’s location is calculated by averaging a set of particles

Each particle represents a random location sample, and has a weight in caculating the object’s location

In each iteration the particle evolves through three stepsPrediction - Predict the location of each particle according to its location at previous time and the object’s moving speed

Updating - Update the weight of each particle according to the sensory data

Resampling – Filter out the particles with small weight.

Existing ApproachesExisting Approaches

Particle Filter (PF)

The accuracy is better than CL/WCL, but continuous

execution of particle filter suffers from high computational

cost

PF achieves high accuracy by sacrificing the computational

efficiency

Our ApproachOur ApproachObservations

WCL are efficienct but the accuracy is not good when the object’s speed is large

Particle filter has satisfactory accuracy but is costly

It is usual that the object moves with a varying speed

Can we integrate the two approaches together to achieve better efficiency as well as accuracy?Hybrid Method

Adaptively switch between using WCL and PF according to the estimated velocity of the moving object

When the speed is low, WCL is used; otherwise, PF is used.

Our ApproachOur Approach

H

tr

Antenna

Tags

d

v

Select the optimal reading range tr

Estimate the object’s speed v

WCL Partical Filter

v < vth v > vth

Technical DetailsTechnical Details

Pratical Issues

How to determine the reading range (or power level) of the

reader?

How to estimate the speed of the moving object?

How to determine the threshold for the speed vth?

Technical DetailsTechnical DetailsHow to adjust the reading range tr (power level) ?Density of tags is presented by its spacing aThe otpimal a/tr is 0.9The necessary tr of the RFID reader should be as large as

1.1a

Technical DetailsTechnical DetailsHow to Estimate the object’s speed?

Calculate the time duration di(N) that each tag i stays in the

reader's reading range in the last N time slots ( )

Select the maximum one dmax(N) = max {di} to estimate the

speed

2tr is the diameter of the reading range; c is scaling constant

depending on the tag density ( 0<c<1 )

The tag with largest di tends to be the one which is closest

to the object's actual moving path

max

2

( )

tr cv

d N

iN d i

Technical DetailsTechnical DetailsConsideration of the threshold vth

Threshold of the speed vth is too low, the hybrid method

has little improvement on the computational cost

compared with particle filter

If vth is set too high, the hybrid method sacrifices the

accuracy too much

We can control the tradeoff between accuracy and

efficiency by chosing proper Nth

2th

th

tr fv

N

EvaluationEvaluation

v (m/s) 0.1 0.2 0.3 0.4 0.5 0.6

WCL

(ms)

119 86 75 64 65 66

PF (ms) 7593 8484 10568 9792 9260 9276

Comparison between WCL and PF

Simulation environment’s size is 4m*4m

The object moves with constant speed along a rectangle

trajectory ( a = 0.27m, f=10Hz, tr = 0.3m )

EvaluationEvaluation

How much does the Hybrid Method improve the performance?

The object's velocity varies during the trajectory

Schemes WCL PF Hybrid

Location

Error (m)

0.187 0.068 0.062

Execution

Times (s)

0.08 4.53 1.26

EvaluationEvaluationIndoor wheelchair tracking

RFID tags are densely deployed in a 4m*6m classroom.

The wheelchair moves along a line with a varying speed,

which is 0.1 m/s at the first half , 0.6m/s at the second half.

UHF RFID reader with a circularly polarized antenna

The reader has 8 power levels, we tune it on level 12.

tr = 60cm,

a = 50 cm,

f = 10Hz,

Nth = 40

EvaluationEvaluationIndoor wheelchair tracking

Experimental Results

The hybrid method almost achieve the same accuracy as

PF, but outperforms PF a lot in term of computational cost

Methods WCL PF Hybrid Method

Location Error (m)

0.482 0.186 0.195

Time per step (ms)

0.2 18.1 7.2

ConclusionsConclusions

We proposed a hybrid method for achieving high accuracy and efficiency in object tracking

The hybrid method is suitable for tracking the mobile object which moves with a varying speed

Our method is demonstrated to be more computational efficient than PF while guaranteing the same accuracy with PF

Thank you!!