These are PPT slides presented for oral presentation corresponding to our regular paper of IPIN 2018.. This presentation was given at the special session "A survey on Indoor Localization Competitions" in IPIN 2018.
Software Delivery At the Speed of AI: Inflectra Invests In AI-Powered Quality
Review of PDR Challenge in Warehouse Picking and Advancing to xDR Challenge
1. National Institute of Advanced Industrial Science and Technology
Review of PDR Challenge in
Warehouse Picking and Advancing to
xDR Challenge (Regular Paper)
Ryosuke Ichikari1, Ryo Shimomura12, Masakatsu Kourogi1,
Takashi Okuma1, and Takeshi Kurata123
AIST1, University of Tsukuba2,
3Sumitomo Electric Industries, Ltd
1
IPIN 2018@Nantes
SS A Survey on Indoor Localization Competitions
10:20 - 12:20 Sept. 26th, 2018
2. National Institute of Advanced Industrial Science and Technology
PDR Benchmarking
• The number of research and development
for PDR is increasing all over the world
• PDR is relative tracking, therefore, the
evaluation method should be different from
one for absolute tracking methods such as
GPS, WiFi based localization
• The format of specification/evaluation in
the papers should be unified for far
comparison
2
Benchmark
Indicators +
Benchmarking
Process
Trial Set
(Dataset)+
3. National Institute of Advanced Industrial Science and Technology
Type of Benchmarking/Competitions
• Off-site benchmarking/competitions
– Organizer prepare shared dataset and evaluate
submitted result by compared with ground truth
e.g.) EvAAL/IPIN competitions (Off-site)
ISMAR tracking competition(Off-site)
KITTI Vision Benchmarking
• On-site benchmarking/competitions
– Organizer prepare controlled environment, participants
gather at the site and run their method.
– Check points are measured by organizer in advance.
e.g.) IPIN competition(On-site), ISMAR tracking
competition(On-site), UbiComp/ISWC PDR Challenge
3
4. National Institute of Advanced Industrial Science and Technology
A Short Survey of Indoor
Localization Competitions
4
5. National Institute of Advanced Industrial Science and Technology 5
Comparison of Indoor Localization Competitions
PerfLoc by NIST EvAAL/IPIN Competitions
Microsoft
Competition@IPSN
Scenario
30 Scenarios
(Emergency scenario)
Smart House/Assisted Living
Competing maximum
accuracy in 2D or 3D
Walking
/motion
walking/running/
backwards/sidestep/
crawling/pushcart/ elevators
(walked by actors on planed
path with CPs)
Walking/Stairs/Lift/Phoning
/Lateral movement (walked by
Actors on planed path with
CPs)
Depends on operators
(developers can operate
their devices by themselves
On-site or
Off-site
Off-site competition and
Live demo
Separated On-site and Off-
site tracks
On-site
Target
Methods
Arm-mounted smartphone
based localization method
(IMU, WiFi, GPS, Cellular)
Off-site: Smartphone base
On-site : Smartphone base/
any body-mounted device
(separated tracks)
2D:Infra-free methods
3D:Allowed to arrange Infra.
(# of anchor and type of devices
are limited on 2018)
# of
people
and trial
1 person × 4 devices
(at the same time)
× 30 scenarios
Depends on year and track
(e.g. 9 trials, 2016T3)
N/A
Time
per trial
Total 16 hours
Depends on year and track
(e.g. 15 mins (2016T1,T2), 2
hours (2016T3))
N/A
Evaluation
metric
SE95
(95% Spherical Error)
75 Percentile Error Mean error
History 1 time (2017-2018)
7 times (2011,2012,2013,
(EvAAL),2014,2015(+ETRI),20
16,2017(EvAAL/IPIN))
5 times
(2014,2015,2016,2017,
2018)
6. National Institute of Advanced Industrial Science and Technology 6
Comparison of PDR Challenges
Ubicomp/ISWC 2015 PDR
Challenge
PDR Challenge in Warehouse
Picking in IPIN 2017
xDR Challenge for
Warehouse Operations
2018
Scenario
Indoor pedestrian
navigation
Picking work inside a
logistics warehouse
(Specific Industrial Scenario)
General warehouse
operations including picking,
shipping and driving forklift
Walking
/motion
Continuous walking while
holding smartphone and
looking at navigation screen
Includes many motions
involved in picking work, not
only walking
Includes many motions
involved in picking, shipping
operations and, not only
walking. Some workers
may drive forklift
On-site or
off-site
Data collection: on-site
Evaluation: off-site
Off-site Off-site
Number
of people
and trial
90 people, 229 trials 8 people, 8 trials
34 people + 6 forklifts,
170 trials (PDR) +
30 trials (VDR)
Time
per trial
A few minutes About 3 hours About 8 hours
Evaluation
metric
Mean Error, SD of Error
Integrated Evaluation
(integrated by accuracy,
naturalness, warehouse
dedicated metrics)
Integrated Evaluation
(integrated by accuracy,
naturalness, warehouse
dedicated metrics)
Remark
Collection of data of
participants walking. The
data are available at HASC
(http://hub.hasc.jp/) as
corpus data
Competition over integrated
position using not only PDR,
but also correction information
such as BLE beacon signal,
picking log (WMS), and maps
Consists of PDR and VDR
tracks.Referential motion
captured by MoCap. also
shared for introducing
typical motions.
7. National Institute of Advanced Industrial Science and Technology
(Original)PDR Challenge
• Held in Ubicomp/ISWC2015 as a competition for
PDR
• Purposes:
– Evaluating various PDR engines
– Evaluating evaluation metrics
– Collecting dataset by participants
• The participants submit
source code of algorithm
based on app skeleton.
7
8. National Institute of Advanced Industrial Science and Technology 8
• Held “PDR Challenge in warehouse picking” as the official
competition track (Track 4) of IPIN (Int. Conf. on Indoor Positioning
& Indoor Navigation) in Sapporo, Japan 2017
• IPIN 2017 competitions: total 4 tracks, 20 teams joined (CN5, KR4,
JP3, TW2, GE2, AU1, FR1, CL1, PT)
• Track4: Off-site PDR competition for tracking workers during picking
operations in actual warehouse
• Chairs: Ichikari & Kourogi from AIST WMS: warehouse management system
PDR Challenge in Warehouse Picking
9. National Institute of Advanced Industrial Science and Technology
How to collect raw sensor data by smartphone
Device:Android Nexus 5
Collected raw sensor data required for
PDR
(Acceleration,Angular velocity,
magnetism,Atmospheric pressure, BLE
signal log from beacons) ~100Hz
Sensor data for Calibration
Measured while fixing for a whole
8 shaped movement for magnetic
calibration
10. National Institute of Advanced Industrial Science and Technology
Shared data for PDR Challenge 2017
10
P1 P2
P3
WMS picking log
(WMS:Warehouse Management System)
Sensor raw data
(gyro. ,acceleration, magnetism,
air pressure, BLE log)
Specification of the target
warehouse
(Map, Shelves, Obstacle)
11. National Institute of Advanced Industrial Science and Technology
Ground truth is extracted from WMS picking log
11
X Y time1
X Y time2
X Y time3
WMS data
X Y time n-2
X Y time n-1
Share with
participants
for error correction
Hided for
participants and
used for evaluation
(as GT)
X Y time n-3
X Y time0
X Y time 1
X Y time 2
X Y time 3
X Y time 0
X Y time n-2
X Y time n-1
X Y time n
X Y time n-3
X Y time n
Original data
12. National Institute of Advanced Industrial Science and Technology
12
Metrics related to accuracy
- Metric related to integrated positioning error
(Ed)
- Metric related to PDR error based on EAG
(Es)
Metrics related to the trajectory naturalness
- Metric related to the naturalness of travel
speed (Ev)
- Metric related to position measurement
output frequency (Ef)
Specific metrics for warehouse picking scenario
- Metric related to collision with obstacles (Eo)
- Metric related to motions during picking
work (Ep)
Evaluation Metric
Comprehensive evolutions(C.E.)
Evaluation Metric
13. National Institute of Advanced Industrial Science and Technology
Metric related to integrated positioning error
(Ed)
13
Pos Err: Vector consists of all positional
error at the ground truth available points
Calculating positional error as
Euclid distance between
ground truth by WMS and
points in submitted trajectory
14. National Institute of Advanced Industrial Science and Technology
Metric related to PDR accumulating error (Es)
14
EAG
EAG
EAG
EAG
Es is based on speed for error
accumration called
Error Accumulation Gradient (EAG)
calculated by elapsed time from
correcting points and error
Ref.:
15. National Institute of Advanced Industrial Science and Technology
Example of error plot for obtaining EAG
15
In PDR Challenge2017, we adopt simple linear regression whose
intersection equal to 0 for calculating representative EAG
EAG[m/sec]
16. National Institute of Advanced Industrial Science and Technology
Metric related to motions during picking work
(Ep)
16
・Evaluating naturalness during picking operations
・Assumed that worker should stop in front of selves during
the picking
・We check the total length of the movement measured from
1.5 sec before the picking to 1.5 sec after the picking is less
than 1 meter
17. National Institute of Advanced Industrial Science and Technology
Metric related to the naturalness of travel speed
(Ev)
17
・We assumed natural human travel speed is less than 1.5m/sec.
・We check whether local speed is less than 1.5m/sec or not
18. National Institute of Advanced Industrial Science and Technology
Metric related to collision with obstacles (Eo)
18
・We assume that ideal trajectories do not enter the area
where employees cannot walk inside or pass over
・This metric quantify the degree of incursion of the trajectories
into the forbidden area
・We defined a tolerance area with 0.17 m width around
borders of the forbidden area for ignoring small amount of the
incursion.
19. National Institute of Advanced Industrial Science and Technology
Metric related to output frequency (Ef)
19
・Ideal trajectory should frequently submitted with less than
certain length of interval.
⇒ Frequency is less than 1Hz should be deducted
・Local frequency can be calculated by elapsed time from the
previous submitted points in the trajectory
20. National Institute of Advanced Industrial Science and Technology
Example of submitted trajectories
• Terminal 2
20
21. National Institute of Advanced Industrial Science and Technology 21
Results of the evaluation metrics and final C.E.
Result of Metric Eo
Result of EAG for Metric Es
PDR Challenge in Warehouse Picking
Example:EAG:0.12m/sec. ⇒ 7.2m/min,
Target accuracy of the integrated localization: 4.0 m
Guideline of absolute positioning method: every 30 sec. & 0.4 m or less error (3.6+0.4=4.0m)
Team Ed Es Ep Ev Eo Ef Median
of Error
[m]
Median
of EAG
[m/s]
C.E.
Team1 66.876 93.692 97.195 99.998 51.821 11.323 10.606 0.173 68.652
Team2 71.524 94.872 43.545 100 100 9.258 0.150 90.419
Team3 76.459 95.333 72.719 87.835 93.549 99.271 7.827 0.141 89.161
Team4 51.934 90.769 84.965 95.657 59.623 99.239 14.939 0.230 74.948
Team5 78.386 96.308 97.484 99.093 45.530 100 7.268 0.122 78.336
AIST 80.272 96.718 81.057 98.711 89.968 95.879 6.721 0.114 90.836
99.876
22. National Institute of Advanced Industrial Science and Technology
Findings from the PDR Challenge 2017
• We successfully encouraged the competitors to
develop practical localization methods for warehouse.
• Are the weights for evaluation metrics appropriate?
– The weighs for obstacle interference seems to be too high.
– Weights for error evaluation should be added.
• There are not much difference in PDR error
evaluation metric Es
– Is linear regression for calculating EAG is appropriate?
• Workers who drive forklift cannot be tracked.
• The competitors (researchers) are not familiar with
the warehouse operations.
• Scale should be extended.
23. National Institute of Advanced Industrial Science and Technology
Re-consideration of the weights (1)
• Results with original weights at the competition
Ed(median_error):Es(EAG):Ep(picking):Ev(velocity):Eo(obstacle):Ef(frequency)
=20%:20%:5%:15%:30%:10%
23
Team Ed Es Ep Ev Eo Ef Median
Error
[m]
Median
EAG
[m/s]
C.E.
Team1 66.876 93.692 97.195 99.998 51.821 11.323 10.606 0.173 68.652
Team2 71.524 94.872 43.545 100 99.876 100 9.258 0.150 90.419
Team3 76.459 95.333 72.719 87.835 93.549 99.271 7.827 0.141 89.161
Team4 51.934 90.769 84.965 95.657 59.623 99.239 14.939 0.230 74.948
Team5 78.386 96.308 97.484 99.093 45.530 100 7.268 0.122 78.336
24. National Institute of Advanced Industrial Science and Technology
Re-consideration of the weights (2)
• Trial by adding weight for error metrics (Es(median_error),
Ed(EAG)) and reducing weight of Eo(obstacle).
• The weights will be replaced with updated one.
24
Ed(median_error):Es(EAG):Ep(picking):Ev(velocity):Eo(obstacle):Ef(frequency)
=30%:30%:5%:10%:15%:10%
Team Ed Es Ep Ev Eo Ef 50%
eCDF
[m]
50%
eCDF
[m/s]
C.E.
Team1 66.876 87.053 97.195 99.998 51.821 11.323 10.606 0.173 69.944
Team2 71.524 89.474 43.545 100 99.876 100 9.258 0.150 85.458
Team3 76.459 90.421 72.719 87.835 93.549 99.271 7.827 0.141 86.443
Team4 51.934 81.053 84.965 95.657 59.623 99.239 14.939 0.230 72.577
Team5 78.386 96.308 97.484 99.093 45.530 100 7.268 0.122 84.021
25. National Institute of Advanced Industrial Science and Technology
xDR Challenge for Warehouse Operations
• xDR Challenge for Warehouse Operations 2018 was held as an
sequel competition to "PDR Challenge in Warehouse Picking“
• Host:PDR Benchmark Standardization Committee
• Competition of Dead-Reckoning for Pedestrian and Vehicle
– xDR=PDR+VDR
• Important dates
– Testing period: Mid May to Mid Aug., 2018
– Results submission due: 18th Sept, 2018
– Announcement of Winners:
Now (This Special Session)
• Sponsors:
25
26. National Institute of Advanced Industrial Science and Technology
Two competition tracks
• Individually determining winners in each tracks
• PDR-Track
– Tracking workers who move by foot during warehouse
operations
– Shared data: Smartphone sensor data for PDR, BLE tag’s
signal, warehouse’s spec, Partial WMS log. reference movie
for typical picking
• VDR-Track
– Tracking forklift driven by employee during warehouse
operations
– Smartphone sensor data measured by attaching
smartphone onto the forklifts
– Shared data: Smartphone sensor data for VDR, BLE tag’s
signal, warehouse’s spec, forklift spec., partial WMS log.
– Sample data with known path are prepared. (For beginners)
26
27. National Institute of Advanced Industrial Science and Technology 2727
Ubicomp/ISWC 2015 PDR
Challenge
PDR Challenge in Warehouse
Picking in IPIN 2017
xDR Challenge for
Warehouse Operations
2018
Scenario
Indoor pedestrian
navigation
Picking work inside a
logistics warehouse
(Specific Industrial Scenario)
General warehouse
operations including picking,
shipping and driving forklift
Walking
/motion
Continuous walking while
holding smartphone and
looking at navigation screen
Includes many motions
involved in picking work, not
only walking
Includes many motions
involved in picking, shipping
operations and, not only
walking. Some workers
may drive forklift
On-site or
off-site
Data collection: on-site
Evaluation: off-site
Off-site Off-site
Number
of people
and trial
90 people, 229 trials 8 people, 8 trials
34 people + 6 forklifts,
170 trials (PDR) +
30 trials (VDR)
Time
per trial
A few minutes About 3 hours About 8 hours
Evaluation
metric
Mean Error, SD of Error
Integrated Evaluation
(integrated by accuracy,
naturalness, warehouse
dedicated metrics)
Integrated Evaluation
(integrated by accuracy,
naturalness, warehouse
dedicated metrics)
Remark
Collection of data of
participants walking. The
data are available at HASC
(http://hub.hasc.jp/) as
corpus data
Competition over integrated
position using not only PDR,
but also correction information
such as BLE beacon signal,
picking log (WMS), and maps
Consists of PDR and VDR
tracks.Referential motion
captured by MoCap. also
shared for introducing
typical motions.
Comparison of PDR Challenges
28. National Institute of Advanced Industrial Science and Technology
Prizes
• VDR Track: (a) { VDR module (SSEI, Eq. to 200,000) +
Android IoT device BL-02 (BIGLOBE) + 150,000 cash}
or (b) {200,000 cash +BL-02}
• PDR Track: (a) {TECCO (Eq. to 100,000) + BL-02 +
150,000 cash}
or (b) {200,000 cash + BL-02}
• Runner-Up:BL-02 + 100,000 cash
28
VDR module TECCO BL-02
29. National Institute of Advanced Industrial Science and Technology
VDR Module (SUC-VDR100)
• Relative vehicle tracking module by VDR
• Manufactured by Sugihara SEI, and its vibration-
based VDR algorithm is licensed by AIST
• Spec
– Battery life:
12 hours
29
30. National Institute of Advanced Industrial Science and Technology
Tecco (TC-A01)
• Wearable RFID-tag reader for picking operation
• Manufactured by GOV
• Spec
– Interface: Bluetooth
– Battery life: 40 hours
30
31. National Institute of Advanced Industrial Science and Technology
Android IoT device (BL-02)
• Android IoT device sold by BIGLOBE
• Ideal characteristics for industrial use
– LTE capable
– No camera (for security/confidential point of view)
– 10-axis sensors for PDR
– Android version (6.0) is fixed.
31
32. National Institute of Advanced Industrial Science and Technology
Rigorous evaluation of error accumulation by BUP
(BLE Unreachable Period)
• Intentionally deleting partial BLE signal logs from the
test data for evaluating PDR accumulated error
Period when BLE signals are deleted: BLE unreachable period (BUP)
• WMS Reference points provided before and after BUP
BUPBUP BUPRSSI
of BLE tag.
Evaluation Points by WMS
⇒ Position data are hided
Correction Points by WMS
⇒ Position data are provided
t
Evaluating positional errors of integrated localization system with BLE beacon
Evaluating accumulated errors caused by only PDR
Emedian_error
Eaccum_error
33. National Institute of Advanced Industrial Science and Technology
Conclusions
• PDR Challenge Advancing to xDR Challenge.
• We promoted and encouraged, but the starting development of VDR
is too challenging.
• The dataset of PDR Challenge & xDR Challenge are very realistic and
good for evaluating practical performance
– Real data
– Huge data including dirty data.
• Survey of the exiting competitions
• Future Works
– Competition whose targets seamlessly changing walking and driving
forklifts (Universal Dead-reckoning: UDR)
– Dead-reckoning for other targets (e.g. Drone Dead-reckoning: DDR)
– Utilizing whole body motion
– Competition of action recognition
33
34. National Institute of Advanced Industrial Science and Technology
Thank you!
• Contact Info.
– Ryosuke Ichikari, Ph.D.(r.ichikari@aist.go.jp)
34
Editor's Notes
This presentation contain the presentation about the regular paper and survey of the exiting competitions and the announcement of
Winer of our new competition xDR challenge.
As you all know the number of R&D about PDR is increasing all over the world.
So you have many choices when you start services.
We need special evaluating method for fair comparison.
We introduce our activity based on benchmarking and competitions.
Basically the element of the benchmarking are these there.
In this presentation, we introduce our benchmarking process and indicators, and dataset.
Benchmarking and competitions have been held in various research area, such as computer vision and indoor localization.
The competitions can be separated into two types:
One type is Off-site competition:
In the off-site competition, …
Off-site tracks of the EvAAL/IPIN competition, and ISMAR competitions are categorized as type.
The other type is on-site competitions.
In this type of competitions …
On-site tracks of the EvAAL/IPIN competition, and ISMAR competitions, and the original PDR Challenge are categorized as type.
Before introducing the our PDR Challenge and xDR challenge,
Let me summarize an existing competitions as a short survey of indoor localization competitions.
In this table, we picked up PerfLoc competitions and EvAAL/IPIN competitions,
As well as the Microsoft competition at IPSN conference,
Which is the other famous competition in the indoor localization,
We summarize competitions in terms of competition’s scenario, including walking style and motion, competitions category,
, target methods, # of people and trials are measured for sharing dataset, and length of the dataset,
,main evaluation metric and its history.
I marked remarkable characteristics as red.
PerfLoc’s summaries are:
There are 30 scenarios for dataset, and its main scenario is localization for emergency scenario
In the variety of scenario, they include various type of walking and motion such as …
The main caricaturists is that they mounted smartphone onto arm, 4 devices at the same time.
Maybe this is good for emergency scenario, but not be good for the general daily scenario.
As the evaluation metric, they adopt 95 % spherical error.
Summery of EvAAL/IPIN competitions are follows:
EvAAL competitions have been held since 2011
It is an one of the oldest competitions,
Their main scenario is smart house and assistive living.
Their competition is not single track, off-site tracks shared smartphone data base,
One on-site track only allow the participant to use smartphone, the other on-site track they allow the participants to any body-mounted devices.
Their main evaluation metrics are 75 percentile error.
Microsoft Competition also has long history and gathered many participants.
Their competition site is relatively small area
So it think they are not based on the realistic scenario.
Therefore, I think Microsoft competition is a competition where the participants compete maximum accuracy on site.
Walking and movement are depending on operators, because thy can operate by themselves
In the Microsoft competition, there are two type of competition track,
2D track, the competitor cannot arrange any infra-structure, 3D track they are allowed to arrange
Infra-structure such as UWB, infra-red.
Here is a competition of PDR Challenge.
All PDR challenges are designed with PDR benchmark standardization committee.
PDR challenges are held under specific scenarios.
Original PDR challenge was held in 2015, the its scenario was
Indoor pedestrian navigation.
PDR challenge in warehouse picking follow the concepts.
Its’ scenario is industries scenario, especially warehouse.
We started with picking operation first year, and We added general warehouse operation including driving forklift this year.
This slide shows the introduction of original PDR challenge littlie bit in detail.
The original PDR challenge was held in UbiComp/ISWC2015 as a completions for PDR.
The purpose of the competitions are:
-
-
Actually there was a separated track for presenting the proposal of the evaluation method.
-
The participant are required to submit source code of the algorithm.
They wrote based on the app skeleton provided by the organizer.
We held a tracking competition called PDR challenge in warehouse picking as an official competition track of IPIN last year.
Our track was track 4, it was an off-site competition track
In our, the competitors compete performance of the indoor localization regarding various aspect using the data measured during actual picking operation in real warehouse.
The chairs of the track are my self and dr. Kourogi from AIST.
This slide shows how to collect the raw sensor data by smartphone.
We asked to the worker to wear the pouch with smartphone putted in it during the work.
We adopt Nexus 5, and measure acceleration, angler velocity, magnetism, atmospheric pressure, and BLE signal log from beacon.
We captured some calibration data for them.
Here are all data shared with competitors in PDR Challenge 2017
In addition to the sensor data,
We shared picking log from the warehouse management system log
It is used for managing items in the selves for daily operation.
Importance thing is the picking log can be converted into record of the workers position at
The specific moment.
We shared detailed specification of the target warehouse as well.
In our competitions, the ground truth used for evaluating submitted results is extracted from the WMS picking log.
Some part of the ground truth data are shared for competitors for error correction.
And others are hided for participants, they are used for error evolution
Evaluation metric is one of important part for our PDR Challenge.
Most of the competition only evaluate the error compared with ground truth.
In addition to the error, we evaluate practical performance of the method with multi-faceted evaluation metrics
For the metrics related accuracy
(省略)
We evaluated integrated positioning error Ed
We evaluated PDR accumulating error based on EAG as Es
For metrics to evaluating naturalness of the trajectory
We evaluate the naturalness of speed as Ev
We evaluate the frequency points in the trajectory as Ef
For specific evaluation metric for warehouse scenario.
- We evaluate collision with the obstacle in the warehouse.
Let me introduce detail of each metric
For calculating the integrated position error Ed,
We adopt following formula.
In this, we calculate positional error as Euclid distance between ground truth by WMS
And submitted trajectory.
Next metric for accuracy is PDR accumulation error metric Es
We assume that the participants can cancel the error to 0 at the reference points and
error is increased as time passes
Es is based on ・・・
Here is an example of error plot for EAG
The EAG can be calculated every reference points.
Therefore we can get the eCDF (Empirical cumulative distribution function) of EAG.
We can determine the repesentativ EAG by eCDF such as median or 75th percentile EAG.
In PDR challenge ・・・
Next metric is related to motions during the picking work. (Ep)
With this metric we evaluate ・・・
We assume.
We check.
Here is the detailed calculation for the metric related to naturalness of the speed.
We assume
We check --- with following formula.
…
Here is the detailed calculation for the metric related to to collision with obstacle
This is belong to the warehouse dedicated scenario, but can be applied for other cased as well.
We assume
This metric
We define …
…
Here is the detailed calculation for the metric related to frequency
Ideal trajectory ・・・
This metric encourage the method to submit as many as possible,
But just points with high confidence.
Finally, I will introduce the result of the PDR Challenge 2017.
Here are submitted trajectory,
As you can see, the trajectories from each competitors look different.
Here is final results evaluated by our evaluation metrics.
In our track we got five competitors from 4 country. And added our result just for reference
The team 2 won the competition, and their algorithm is tuned very well for our regulation.
One of most interesting finding is error accumulation gradient as introduced in previous slide.
For example, given, the EAG is 0.12m/sec and target accuracy of the localization 4.0 m.
We can get guideline for the configuration of the system from the EAG
If the PDR can be corrected by high quality absolute positioning method with less than 0.4m error every 30 second.
Then we can achieved target accuracy. According to the guideline, we can arrange beacon or other positioning method
Consequently, we can get many participants and findings. We believed we have successfully organized the competition.
Here is a summery of findings and review of the PDR Challenge 2017.
We successfully
・・・ by adopting multi-faceted evaluation metric.
But The detail of the evaluation metrics are open to question.
First question is Are the weights for evaluation metrics appropriate?
The weighs for obstacle interference seems to be high.
A team tuned for this metric very well, and trajectory is not natural as the PDR trajectory
Therefore, the Weights for error evaluation should be added.
Another unexpected fact is there are not much difference in PDR error evaluation metric Es.
We need to further investigate linear regression is appropriate for representative EAG or not.
Other reviews for improving the competitions are
Many forklift are operated for carrying big and heavy items or picking up from high selves.
Workers who drive forklift cannot be tracked only from the PDR
The competitors (researchers) are not familiar with the warehouse operations.
We need to have a way to introduce what the picking operation is
Scale of the measurement was not so big.
We reconsider about wights as the internal trial
Here are original weight used in the competitions.
Again, the weight for obstacle interference is 30%, which is very high.
But the some competitors might focus on this one, while discounting the error metrics
This is a kind good strategy for the last year’s competition.
But not always good in general.
This slide show the trial for adding weights for error related metrics Es, Ed,
And reducing the weight of Eo.
The winner is changed, but not so big changeover.
We guess this weights are more suitable for general cases.
In the xDR Challenge, the weight are replaced with these one.
This years competition xDR challenge in warehouse operations 2018 was held as an sequent competition to the PDR challenge.
In the new competition, we added dead reckoning for forklift as the tracking target.
We renamed competitions name xDR challenge: xDR means PDR plus VDR (Vehicle dead-reckoning)
Maybe we are going to add other types of dead-rekoning in the future
Here are important date for the competition, it was very tight schedule for the competitors.
In this presentation we will announce the winner of the competitions.
Our sponsors are: BIGLOBE, Sugihara software & electron industry, GOV, sumitomo electric
Industries, and PDR benchmark Standardization committee.
As I mentioned, In this year, There are two competition tracks
And they individually determine winner in each track.
In the PDR, as the same as last year’s competition,
The competitors’ PDR algorithm is suppose to track worker who move by foot during warehouse operations
This is a off-site competition,
We share the smartphone and warehouse data as same as last year
In the VDR track, which is a new track this year,
The VDR algorithm is supposed to track forklift driven by employee during the operation.
The shared data are saved in same format and contends are almost same with PDR track,
The smartphone data for VDR are measured by attaching smartphone onto the forklift.
For the VDR beginners, we provide sample easy test data with know path.
Also we plan to give some points for those who can only estimate partial element such as speed in operation or not.
Finally, we introduced xDR Challenge parts
The scenario of the xDR Challenge is warehouse work,
It become more general including tracking the forklift.
The added the scale of the data measument,
We measured the warehouse work for 5 Full-business days.
Total 34 people, 6 forklift equals 170 (PDR data ) and 30 VDR data.
About 8 hours / data.
Here are Prizes (読まない)
We awarded winner and runner-up for each tracks.
The winners can cases and extra prize shown in the figure.
One extra prize is VDR module.
As you can see, winner can get the devices which can track the forklift.
As we documented in the regulation documents,
This year, we adopt rigorous evaluation of error accumulation.
Last year, we only evaluation the submitted result which potentially include effort with
PDR, BLE-beacon, WMS, and MAP Map matching. It is hard to extract the effect purely from the PDR.
This year, we Intentionally delete partial BLE signal logs from the test data for evaluating PDR accumulated error.
We call these period the BLE signal are deleted as BLE unreachable period: BUP in short.
With BUP, we evaluation absolute median error of integrated localization only outsize of BUP.
And we evaluate the PDR accumulating error in BUP