1. Lei Yang, Yekui Chen, Xiang-Yang Li, Chaowei Xiao, Mo Li, Yunhao Liu
Tagoram: Real-Time Tracking of Mobile RFID Tags to High-Precision Accuracy Using COTS Devices
Xiang-Yang Li
Lei Yang, Yekui Chen, Xiang-Yang Li,
Chaowei Xiao, Mo Li, Yunhao Liu
Sep. 9, 2014
Good afternoon.
Today I will talk about

2. Lei Yang, Yekui Chen, Xiang-Yang Li, Chaowei Xiao, Mo Li, Yunhao Liu
Tagoram: Real-Time Tracking of Mobile RFID Tags to High-Precision Accuracy Using COTS Devices
Xiang-Yang Li
Lei Yang, Yekui Chen, Xiang-Yang Li,
Chaowei Xiao, Mo Li, Yunhao Liu
Sep. 9, 2014
The challenge to achieve our goal is due to requirement of

3. 01. 02. 03. 04. 05. 06. 07. 08. Outline Motivation State-of-the-art
Overview
03.
Movement with known track
04.
05.
Movement with unknown track
Implementation & Evaluation
06.
This talk has 8 components, although I will not have time to discuss some of the components in detail.
Pilot Study
07.
Conclusion
08.

4. 1
Motivation

5. Imagine you can localize RFIDs to within 0.1cm to 1 cm!
Imagine that …..
This enables many interesting applications.
For example, it enables some innovative human-computer interface.
It allows us to draw, and move in the space freely to control the smart devices
Human-computer interface

6. Imagine you can localize RFIDs to within 0.1cm to 1 cm!
We can also improve the baggage sortation efficiency in airport. More details will be introduced in our pilot study.
Baggage sortation in airport

7. Imagine you can localize RFIDs to within 0.1cm to 1 cm!
Another possible example is Goal-line technology.
It could help to determine when the ball has slightly crossed the goal line. We agree that it is very challenging to do it using RFID.
Hawk-Eye is an existing technology using cv, which require 30% exposure of ball.
GoalRef and Cairos do not have good accuracy.
Goal-line technology

8. Demonstration http://young.tagsys.org/tracking/tagoram/youtube
Before presenting our technical details, let us see a short demo on what we can achieve.

9. High-Precision RFID Tracking Using COTS Devices
Drawing in the Air
40cm
First is about drawing and controlling using RFID.
We showed that a user can draw ANY shape in a two-dimensional space (or 3-D space).
Here are some examples. The left part is the ground-truth figure where a user will follow to draw
and the right one is the one recovered when user follows the pattern to draw.
The mean accuracy is surprisingly good. About 1.2 cm.
More will be shown later in demo session.
So please check out demo and try to draw something yourself.
30cm

10. TrackPoint Deployed at Airports
Reader
The second application of our system is airport baggage sortation.
We deploy our product, TrackPoint, for baggage sortation in two airports already.
To track the location of baggages, and help workers to sort them correctly and efficiently,
we design and customize a tracking device, called as TrackPoint.
It composes of one RFID reader and three RF antennas, placed across the conveyor.
It can localize the baggage to accuracy of about 6cm in real system.
Industrial computer

11. 2
State-of-the-art
Many RFID localization methods have been proposed in the literature.

12. RFID Tracking & Localization
State-of-the-art
[1] Lionel M. Ni, Yunhao Liu, etal
LANDMARC: Indoor Location Sensing Using Active RFID
Wireless Networks
1120mm
LANDMARC Lionel M. Ni
LANDMARC: The RSS of reference tags deployed at known positions are recorded and used to locate the target tags.
2004
2010
2011
2013
2014

13. RFID Tracking & Localization
State-of-the-art
LANDMARC Lionel M. Ni
[1] Lionel M. Ni, Yunhao Liu, etal
LANDMARC: Indoor Location Sensing Using Active RFID
Wireless Networks
1120mm
[3] P.Nikitin, etal
Phase based spatial identification of uhf rfid tags.
IEEE RFID 2010.
680mm
[2] C.Hekimian-Williams, elta
Accurate localization of rfid tags using phase difference.
280mm
Diff based
C. H.Williams
Spatial based
P. Nikitin
2004
2010
2011
2013
2014

14. RFID Tracking & Localization
State-of-the-art
[4] Salah Azzouzi, etal
New measurement results for the localization of uhf RFID transponders using an angle of arrival (aoa) approach
IEEE RFID 2011.
300mm
210mm
[5] R. Miesen etal.
Holographic localization of passive uhf rfid transponders.
AOA based
S. Azzouzi
Holographic
R. Miesen
[1] Lionel M. Ni, Yunhao Liu, etal
LANDMARC: Indoor Location Sensing Using Active RFID
Wireless Networks
1120mm
Diff based
C. H.Williams
Spatial based
P. Nikitin
[3] P.Nikitin, etal
Phase based spatial identification of uhf rfid tags.
IEEE RFID 2010.
680mm
[2] C.Hekimian-Williams, elta
Accurate localization of rfid tags using phase difference.
280mm
2004
2010
2011
2013
2014

15. RFID Tracking & Localization
State-of-the-art
AOA based
S. Azzouzi
Holographic
R. Miesen
[4] Salah Azzouzi, etal
New measurement results for the localization of uhf RFID transponders using an angle of arrival (aoa) approach
IEEE RFID 2011.
300mm
210mm
[5] R. Miesen etal.
Holographic localization of passive uhf rfid transponders.
[1] Lionel M. Ni, Yunhao Liu, etal
LANDMARC: Indoor Location Sensing Using Active RFID
Wireless Networks
1120mm
680mm
[2] C.Hekimian-Williams, elta
Accurate localization of rfid tags using phase difference.
IEEE RFID 2010.
280mm
[3] P.Nikitin, etal
Phase based spatial identification of uhf rfid tags.
[6] Jue Wang, etal
Dude, where's my card?: RFID positioning that works with multipath and non-line of sight
ACM SIGCOMM 2013
110mm
12.8mm
[7] Jue Wang, etal
RF-compass: robot object manipulation using RFIDs
ACM MobiCom 2013
PinIt
Jue Wang
RF-Compass
2004
2010
2011
2013
2014

16. RFID Tracking & Localization
State-of-the-art
[4] Salah Azzouzi, etal
New measurement results for the localization of uhf RFID transponders using an angle of arrival (aoa) approach
IEEE RFID 2011.
300mm
210mm
[5] R. Miesen etal.
Holographic localization of passive uhf rfid transponders.
[1] Lionel M. Ni, Yunhao Liu, etal
LANDMARC: Indoor Location Sensing Using Active RFID
Wireless Networks
1120mm
[8] Tianci Liu, etal,
Anchor-free Backscatter Positioning for RFID Tags with High Accuracy
IEEE INFOCOM 2014
128mm
[9] Jue Wang, etal,
RF-Idaw: Virtual Touch Screen in the Air Using RF Signals
37mm
BackPos
Tianci Liu
RF-IDRaw
Jue Wang
[3] P.Nikitin, etal
Phase based spatial identification of uhf rfid tags.
IEEE RFID 2010.
680mm
[2] C.Hekimian-Williams, elta
Accurate localization of rfid tags using phase difference.
280mm
PinIt
Jue Wang
RF-Compass
[6] Jue Wang, etal
Dude, where's my card?: RFID positioning that works with multipath and non-line of sight
ACM SIGCOMM 2013
110mm
27.6mm
[7] Jue Wang, etal
RF-compass: robot object manipulation using RFIDs
ACM MobiCom 2013
2004
2010
2011
2013
2014

17. RFID Tracking & Localization
State-of-the-art
[4] Salah Azzouzi, etal
New measurement results for the localization of uhf RFID transponders using an angle of arrival (aoa) approach
IEEE RFID 2011.
300mm
210mm
[5] R. Miesen etal.
Holographic localization of passive uhf rfid transponders.
[1] Lionel M. Ni, Yunhao Liu, etal
LANDMARC: Indoor Location Sensing Using Active RFID
Wireless Networks
1120mm
BackPos
Tianci Liu
RF-IDRaw
Jue Wang
[8] Tianci Liu, etal,
Anchor-free Backscatter Positioning for RFID Tags with High Accuracy
IEEE INFOCOM 2014
128mm
[9] Jue Wang, etal,
RF-Idaw: Virtual Touch Screen in the Air Using RF Signals
37mm
[3] P.Nikitin, etal
Phase based spatial identification of uhf rfid tags.
IEEE RFID 2010.
680mm
[2] C.Hekimian-Williams, elta
Accurate localization of rfid tags using phase difference.
280mm
[6] Jue Wang, etal
Dude, where's my card?: RFID positioning that works with multipath and non-line of sight
ACM SIGCOMM 2013
110mm
27.6mm
[7] Jue Wang, etal
RF-compass: robot object manipulation using RFIDs
ACM MobiCom 2013
WE ARE HERE
7mm
2004
2010
2011
2013
2014

18. RSS is not a reliable location indicator especially for UHF tags
State-of-the-art Techniques
RSS based Methods 1
RSS
Orientation VS
These previous work can be grouped into several categories.
I only discuss two here.
The first is signal strength based, which has been widely used for localization in wireless networks.
However, it does not provide a reliable location indicator for UHF tags,
because RSS depends on tag’s orientation and antenna gain.
Worse, the orientation changes frequently in mobile environment.
RSS is not a reliable location indicator
especially for UHF tags

19. State-of-the-art Techniques Needs to deploy dense reference tags
Phase based Methods 2
Another one is to exploit the phase information to locate tags.
Representative examples are PinIt and RF-Compass systems.
AOA locates the tag by measuring the phase difference between the received signals at different antennas, the major challenges for these methods are to deal with NLOS.
PinIt or RF-Compas, one of SAR, needs to deploy dense reference tags in advance.
PinIt (SIGCOMM 2013)
RF-Compass (MobiCom 2013)
Needs to deploy dense reference tags

20. Need mm-level localization accuracy achieved
Summary of Challenges
Need mm-level localization accuracy achieved
especially for mobile tags.
Small overhead, COTS devices
infeasible for using many references for a tracking system spanning a long pipeline.
Existing RFID localization techniques cannot be directly applied due to the following reasons.
First, no reported approach can achieve mm-level localization accuracy, especially for mobile tags.
Second, many RFID localization approaches rely on pre-deployed reference tags for accurate calibration,
which is infeasible for a tracking system spanning a long pipeline (e.g., the sortation line in the air- port).
Third, the mobile tag experiences fast-changing environment with non-ideal communication conditions such as
multipath reflection of RF signals, varied orientation of tags, etc.,
which fail most existing localization techniques for their assumption on static communication environment.
Fast-changing environment
multipath reflection of RF signals
varied orientation of tags
Doppler effect

21. 3 How Tagoram works? Overview the basic idea
So, how does our Tagoram work?

22. Backscatter Communication
❷ Double distance
❸ Continuous wave
❶ 𝑩𝒂𝒕𝒕𝒆𝒓𝒚 𝒇𝒓𝒆𝒆
RFID system utilizes the backscatter communication. In the system,
The tags, no battery equipped, purely harvest energy from the reader’s signal.
The signals traverses a total distance of 2𝑑 back and forth between reader and tag.
The reader continuously emit the signals to supply enough energy for tags. It also need to resolve the week backscatter signals from its coninous wave.
Tag’s reflection characteristics, and receiver circuits will all introduce some additional phase rations (we will discuss more details later.).
❹ Device diversity

23. ≈𝟎.𝟎𝟑𝟖𝒎𝒎 accuracy COTS RFID Reader 0.0015 radians (4096 bits)
The resolution of RF phase, supported by COTS reader, can be achieved radians in theory, thereby it can offer about 0.038mm ranging resolution.
This extremely high resolution offers an opportunity to achieve high precision tracking.
𝑰𝒎𝒑𝒊𝒏𝒋 𝑹𝒆𝒂𝒅𝒆𝒓

24. 𝑀×𝑁 Problem definition
{ 𝜃 1,1 , 𝑡 1,1 ,( 𝜃 1,2 , 𝑡 1,2 )⋯,( 𝜃 1,𝑁 , 𝑡 1,𝑁 )}
{ 𝜃 2,1 , 𝑡 2,1 ,( 𝜃 2,2 , 𝑡 2,2 )⋯,( 𝜃 2,𝑁 , 𝑡 2,𝑁 )}
{ 𝜃 𝑀,1 , 𝑡 𝑀,1 ,( 𝜃 𝑀,2 , 𝑡 𝑀,2 )⋯,( 𝜃 𝑀,𝑁 , 𝑡 𝑀,𝑁 )} ⋮
𝑀×𝑁
Antennas & grids:
The tags move within a survellience region monitored by 𝑀 RF antennas.
The surveillance plane is divided into grid, with 𝑊 rows and 𝐿 columns.
2. When the tags moves, the each antenna gets a glimpse of tag’s locations through the phase measurement. In a very short interval, the reader collects M times N phase values.
2. Our problem is that given antenna’s locations, phase measurements, and time stamps, how to find the tag’s coordinates at an arbitrary interrogated time? Here 𝑓 𝑡 is the time-dependent function outputting the tag’s coordinate at time 𝑡.
Surveillance region
Reader antenna
Utilizing antennas’ locations, sampled phase values and timestamps to find out the tag’s trajectory 𝒇(𝒕)?

25. Controllable Case Uncontrollable Case Tagoram Case 1. Case 2.
In this paper, we propose a holistic system, Tagoram, to address the instant tag tracking problem.
Tagoram studies two complementary cases.
Case 1 controllable case, it mostly occurs at a conveyor belt or assembly line where the objects moves on a predefined track with constant speed.
Case 2, uncontrollable case. RFID tags move along irregular and unpredictable trajectory

26. 4 Movement with Known Track Controllable case
Our goal is to find the tag’s trajectory with a known track.
First consider the controllable case.
Our goal is to find the tag’s trajectory with a known track.

27. Inverse Synthetic Aperture Radar
Virtual Antenna Matrix
Inverse Synthetic Aperture Radar
𝐴 1
𝐴 1,1
𝐴 1,2
𝐴 1,3
𝐴 1,4
𝐴 2,1
𝐴 2,2
𝐴 2,3
𝐴 2,4
Conveyor
Initial position 𝑉
𝒇( 𝒕 𝟎 )
We adopt the basic idea of Inverse Synthetic Aperture Radar.
When a tag is conveyed in the belt, the reader measures the phase values in different locations.
Instead, we think the tag stays at a fixed position f(t0) without any motion while each antenna moves in the opposite direction.
The t0 is the timestamp that the tag is firstly read by the reader.
Both the track function and velocity are known in advance, the virtual antennas’ locations can be inferred, so the sole unknown parameter is the initial positions.
In this way, we transfer tracking problem to locate the initial position.
𝐴 2

28. RF Hologram 𝒙 𝒘,𝒍 RF Hologram Surveillance region 𝑾×𝑳 pixels 𝑾×𝑳 grids
We define the RF hologram to indicate the initial positions.
RF Hologram is a likelihood exhibition, using an image to display how a partitioned grid in tag’s motion plane is likely to be the initial position.
In image, each pixel corresponds to a grid in the surveillance plane. The pixel value is the likelihood that the tag locates at the corresponding grid.
RF Hologram
𝑾×𝑳 pixels
Surveillance region
𝑾×𝑳 grids

29. How to define the likelihood?
The key is
How to define the likelihood?

30. RF Hologram Naïve Hologram Augmented Hologram Thermal noise
In this paper, we progressively propose three kinds of holograms: naïve hologram , Augmented hologram, and Differential Augmental hologram.
We will discuss how they deal with two important influential factors: thermal noise and device diversity.
Device diversity
Differential
Augmented Hologram

31. The theoretical signal A virtual interfered signal
Naïve Hologram 𝐴 𝑇
𝑒 𝑱𝜃
The real signal 𝐴 𝑋
𝑒 𝑱ℎ(𝐴,𝑋)
The theoretical signal
𝑒 𝐽𝜃 is the complex representation of a signal.
1. The first curve is the real signal.
2. The second curve is the theoretical signal for location X. where the H(A,X) is the theoretical phase value between antenna A and location X.
3. We construct a virtual interfered signal by subtracting the real signals from the theoretical signals. 𝐴 𝑋
𝑒 𝑱(ℎ 𝐴,𝑋 −𝜃)
A virtual interfered signal

32. Virtual interfered signal
Naïve Hologram
Sum of all signals
Virtual interfered signal
For naïve hologram, the likelihood is defined as the summed amplitude of all constructed virtual interfered waves.
A key point here is that all measured phase values are viewed to originate from the same initial position as we use a virtual antenna matrix.

33. If 𝑿 is the initial location, the waves add up constructively.
Naïve Hologram
Real axis
Imaginary axis
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟏,𝟏 − 𝜽 𝟏,𝟏 )
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟏,𝟐 − 𝜽 𝟏,𝟐 )
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟒,𝟒 − 𝜽 𝟒,𝟒 )
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟒,𝟑 − 𝜽 𝟒,𝟑 )
Each virtual signal is a blue vector with unit amplitude.
If the X is the initial location, all of them will be close to the real axis, and all waves add up constructively.
As a result, the final wave has a very high amplitude.
If 𝑿 is the initial location, the waves add up constructively.

34. If 𝑿 is not the initial location, the waves canceled out.
Naïve Hologram
Imaginary axis
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟏,𝟐 − 𝜽 𝟏,𝟐 )
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟏,𝟏 − 𝜽 𝟏,𝟏 )
Real axis
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟒,𝟑 − 𝜽 𝟒,𝟑 )
On the other hand, if the X is not the initial location, the virtual signal will be randomly distributed in all angles.
These waves will be sort of cancelled out for each other, making the final summed wave with a small amplitude.
𝑒 𝑱(𝒉 𝑿, 𝑨 𝟒,𝟒 − 𝜽 𝟒,𝟒 )
If 𝑿 is not the initial location, the waves canceled out.

35. Naïve Hologram High PSNR
Here we plot an hologram for a linear track movement.
The detailed experiment setup will be introduced later.
We can see that at positions close to the initial location position, the sum grows as the signals constructively add up to each other.
On the other hand, at the positions away from to the initial position,
the signals superimpose at random phase angles and the sum is significantly lower.
The initial position is found out from a large continuous region with high peek signal noise rate. The size of the region determines the tracking error.
A natural question we ask why there is a large region?

36. Influence from Thermal Noise
100 tags
𝟎 ∘ ~𝟒 𝟎 ∘
𝟗𝟐𝟎∼𝟗𝟐𝟔 𝑴𝑯𝒛
−𝟕𝟎∼−𝟑𝟎𝒅𝒃𝒎
Experiment
Observations
Modeling
CDF
We conduct an empirical studies over 100 tags with environment temperature from 0◦ to 40◦C, various frequencies (920 ∼ 926MHz) including 16 channels, and RSS from -70 to -30dbm (different orientations). One set of the results measured at the 5th channel is depicted in the figure. These experiments suggest that the phase measurement results contain random errors, following a typical Gaussian distribution with a standard deviation of 0.1 radians.
𝜃∼𝒩(𝜇,𝜎)
𝐹(𝜃;𝜇,𝜎)
varying with 𝒅
𝝈=𝟎.𝟏

37. How to deal with thermal noise?
Augmented Hologram

38. A probabilistic weight
Augmented Hologram
A probabilistic weight
𝒉 𝑿,𝑨 − 𝜽 ∼𝓝(𝟎, 𝟎.𝟏)
To eliminate the impacts from the thermal noise, we define the augmented hologram.
In naive RF hologram, the amplitudes of signal are uniformly set to one because the measured amplitude, i.e. RSS, is notably distorted as a result of the severe multipath environment.
In AH, we add a probabilistic weight to the each virtual signal.
We define the random variable 𝒉 𝑿,𝑨 − 𝜽. If the grid X is the initial position, the theoretical and measured phase are equal, so difference follows a Gaussian distribution (0,0.1).
We use the probability that signals are emitted from T to A as the weight. In effect, the pixel close to the initial position seems to be augmented. That is why we call it augmented hologram.
Probability of 𝑻→𝑨

39. Augmented Hologram Shift Ground truth Result scattered
We then plot the augmented hologram using the data as displayed in Naïve Hologram.
As expected, the continuous region is now scattered into small parts. The energy is transferred into the center of the red region, making the amplitude of the center point, augmented.
However, there still exists a large discrepancy between the ground truth and the result.
Then what is the reason that leads to this error?

40. Influence from Tag Diversity
At same position
70 tags
100 times
Experiment
Observations
Modeling
Random test
Pass KS-test to be verified over a uniform distribution with 0.5 significant level.
As mentioned in the beginning, the tag’s reflection characteristics, and receiver circuits will all introduce some additional phase rations. So we conduct another experiment to study the tag diversity.
We deploy 70 different tags at the same position in turn. Each tag is interrogated for 100 times and the mean measured phase value is reported.
We observe that the mean measured phase values are distributed among ∼ radians.
Such an observation suggests that the tag’s diversity takes impact on phase measurement and cannot be ignored in practice.
The results Pass KS-test to be verified over a uniform distribution with 0.5 significant level. The tag diversity is hard to infer.
Tag diversity

41. How to eliminate tag diversity?
Differential Augmented Hologram
The next question is how to eliminate the tag diversity?

42. Using the difference of phase values
Differential Augmented Hologram
We revise the definition using the difference of phase values:
each virtual interfered signal will minus that value collected at the first time-slot.
In this way, the tag diversity can be eliminated through the subtraction.
Notice that the tag diversity did appear in each virtual interfered signal
For the details please refer to our paper.
Using the difference of phase values

43. Differential Augmented Hologram
Ground truth
Result
We then plot the effect of differential augmented hologram using the same data.
We can see that the result almost coincides with the ground truth.
In our experiment, DAH can achieve a mm-level accuracy.

44. Movement with
Unknown Track
Uncontrollable case 5

45. Overview Estimating Speeds, Fitting tag’s trajectory1
Selecting the optimal trajectory 2
Tagoram acquires the phase measured by 𝑀 antennas from different directions to fit tag’s trajectory every around of antenna scheduled.
2. Tagoram utilizes the DAH to evaluate all fitted trajectories and select the optimal one corresponding to the pixel with maximum PSNR.

46. 6 Implementation & Evaluation Purely based on COTS devices
We implement our methods and evaluate their performances
purely….

47. Hardware & Software Introduction
ImpinJ R420 reader.
4 directional antennas
Reader
Tag
Alien 2 × 2 Inlay
Alien Squiggle Inlay
Software
EPCglobal LLRP
Java
Here are the specifications of reader, tag and software.
In our system,
Reader:
We adopt an ImpinJ Speedway modeled R420 reader without any hardware or firmware modification. The reader supports
four directional antennas at most, being compatible with EPC
Gen2 standard.
Tag:
Two types of tags from Alien Corp [16], modeled 2 × 2 Inlay and Squiggle Inlay, are employed. Both of them are employed in our lab experiment and pilot study.
Software:
We adopt LLRP protocol [18] to communicate with the reader. ImpinJ reader extends this protocol for supporting the phase report. We adjust the configuration of reader to immediately report reading whenever tag is detected. The software is implemented using Java.

48. Linear Track
A mobile object is emulated via a toy train on which a tag is attached.
A camera is installed above the track to capture the ground truth.
This figure show the experiment setup using linear track.
The system triggers the camera to take a snapshot on the train when firstly read.

49. Tracking Accuracy in Linear Track
DAH 𝟒 𝟏𝟑 𝟏𝟒
RSS: The difference between the RSSs of a pair of tags is used as an indicator for their spatial distance as in past work. The RSS scheme has a combined accuracy of 600mm, suffering from the high variation in behavior across tags, like the antenna gain and tag’s orientation.
Otrack: [3] designed a method to determine baggage order via the changes of RSS and read rate.
PinIt: PinIt [1] uses synthetic aperture radar (SAR) created via antenna motion to extract the multipath profiles for each tag, and leverages the reference tags to locate the target tag as same way as RSS based methods.
Our method DAH has a median accuracy of 5mm, 13mm and 14mm in X-axis, Y -axis and combined dimension, outperforming RSS, OTrack, PinIt and BackPos by 48×, 30×, 28× and 8.5× respectively.
Improve the accuracy by 𝟒𝟖×, 𝟑𝟎×, 𝟐𝟖× 𝒂𝒏𝒅 𝟖.𝟓× in comparison to RSS, OTrack, PinIt and BackPos.

50. Controllable Case with Nonlinear Track
The track’s internal diameter is 540mm We
The distance varies from 𝟏𝒎 to 𝟏𝟓𝒎

51. Nonlinear track
DAH
𝟒.𝟗𝟖
𝟐.𝟗𝟓
𝟕.𝟐𝟗
DAH has a mean error distance of 4.14mm, 4.98mm and 7.29mm in x-axis, y-axis and combined dimension, outperforming others by 82×, 21×, 16× and 55× respectively.
The accuracy of DAH in circular track is better than that of linear track. In this situation, the accuracies along the x-axis and y-axis perform equally well, because the tag takes almost the same displacement along two dimensions.

52. Under Uncontrollable Case
We then study the tacking accuracy in a uncontrollable case where the track function is not known in advance.
The figures illustrate the final fitted trajectory and speed chains. As a result, Tagoram achieves a median of 12.3cm accuracy with a standard deviation of 5cm for moving toy-cars.
Accuracy is significantly improved when moving speed is low.
Achieve a median of 12.3cm accuracy with a standard deviation of 5cm.

53. 7
Pilot Study
In two airports

54. Current workflow – Manual sortation
It is error-prone step to find the baggage from the carousel in manual sortation.
To reduce the cost, tens of thousands of baggages enter conveyor system to be transported by a shared conveyor system every day in an airport.
The baggages from different check-in counters are mixed together.
Often uses human power to separate these baggages according to their flight numbers.
The sorting task is finished in the sortation carousel.
The sorters, each of who is in charge of sorting one flight of baggages, stand by the carousel along the conveyor.
His mission is to find out and carry the target baggages from the carousel to the cart.
It is an error-prone step for the sorter to pick an expected one out of packs of baggages.
Sortation carousel
Sorting baggage

55. Our system: TrackPoint
Reader
The conveyor system comprises a huge mechanical network and could be hundreds of meters long.
It is impossible to deploy hundreds of readers along the conveyor belt for full coverage.
So we design and customize a tracking device, called as TrackPoint.
It composes of one RFID reader and three RF antennas, placed across the conveyor. When a tag is passing through the TrackPoint, it is interrogated repeatedly by the reader and tracked using Tagoram with DAH.
Industrial computer

56. Tracking Visualization
To help improve the sorting accuracy and efficiency, we built four large screens, to show the tracked baggages’ trajectories, assisting
the sorters to find out their baggages.
In the screen, the baggage carousel is proportionally displayed, where there are lots of small
Squares representating baggages.
The baggages from different flights are filled with different colors.
With the help of baggage
visualization and accurately tracking, the sorter can quickly and clearly catch sight of the bags which he is responsible for, and fast
position the baggages when searching them.

57. Pilot site Beijing Capital International Airport
Sanya Phoenix International Airport
We have instrumented our system in two airports, BCIA T1 and SPIA, both of which employ the manual sortation

58. Setup
Each airport contains 5 TrackPoints, 4 visualization screens, and 22 RFID Printers.
The two-year pilot study totally spent more than $600,000

59. Setup Consumed 110,000 RFID tags.
Involved 53 destination airports, 93 air lines, and 1,094 flights.

60. Tagoram achieves a median accuracy of 6.35cm in practice.
Tracking Accuracy
Since the TrackPoint’s positions are unknown and fixed, we measure the tracking accuracy by comparing two consecutive read time for a same baggage between neighbored TrackPoints, The accuracy is shown in figure.
We find Tagoram achieves a median accuracy of 63.5mm.
Tagoram achieves a median accuracy of 6.35cm in practice.

61. Conclusion
Provides real-time tracking of mobile tags with accuracy to mm-level.
Limits almost all negative impacts.
Multipath effect
Doppler effect
Thermal noise
Device diversity
Implemented purely based on COTS RFID products.
Systematically evaluated under indoor environment and at two airports.
Not well-studied before
Existing RFID localization techniques cannot be directly applied due to the following reasons.
First, to the best of our knowledge there is no reported approach that can achieve mm-level localization accuracy, especially for mobile tags. Second, many RFID localization approaches rely on pre-deployed reference tags for accurate calibration, which is infeasible for a tracking system spanning a long pipeline (e.g., the sortation line in the air- port).
Third, the mobile tag experiences fast-changing environment with non-ideal communication conditions such as multipath reflection of RF signals, varied orientation of tags, etc., which fail most existing localization techniques for their assumption on static communication environment.

62. Questions?
hank you T