www.DLR.de • Chart 1 > IPIN 2013 > Patrick Robertson
Simultaneous Localization and Mapping for Pedestrians using Distortions of the Local Magnetic Field Intensity in Large Indoor Environments
Patrick Robertson1, Martin Frassl1, Michael Angermann1, Marek Doniec2, Brian J. Julian2, Maria Garcia Puyol1, Mohammed Khider1, Michael Lichtenstern1, and Luigi Bruno1
1 Institute of Communications and Navigation, German Aerospace Center (DLR)2 Computer Science and Artificial Intelligence Laboratory, MIT
Presented at IPIN 2013,
October 2013, Montbeliard,
France
www.DLR.de • Chart 2 > IPIN 2013 > Patrick Robertson
Magnetic Localization in Buildings?
- Listing- Second Level
- Third Level- Fourth Level
- Fifth Level
Photo source: The Salt Lake Tribune
“One person’s noise
is another’s signal” –
Brad Parkinson
www.DLR.de • Chart 3 > IPIN 2013 > Patrick Robertson
Human Step Measurement: Odometry
www.DLR.de • Chart 4 > IPIN 2013 > Patrick Robertson
- Zero Velocity
Updates during still
phase!
- Pioneering work:
Foxlin, 2005- Proposition: Use a foot mounted IMU with co-
located magnetometer.
“Close to the ground lots of signal”
Unaided Human Odometry Exhibits Drift
www.DLR.de • Chart 5 > IPIN 2013 > Patrick Robertson
Magnetic Localisation in Buildings
- Goldenberg, Geomagnetic navigation beyond the magnetic compass, 2006
- Haverinen, Kemppainen, A global self-localization technique utilizing local
anomalies of the ambient magnetic field, 2009
- Zhang, Martin, Robotic mapping assisted by local magnetic field anomalies, 2011
- Gozick et al, Magnetic maps for indoor navigation, 2011
- Riehle et al., Indoor waypoint navigation via magnetic anomalies, 2011
- Kim et al., Indoor positioning system using geomagnetic anomalies for
smartphones, 2012
- Le Grand, Thrun, 3-axis magnetic field mapping / fusion for indoor localization, 2012
- Angermann et al., Characterizing Magnetic Field in Buildings, 2012
- Frassl et al., Magnetic maps of indoor environments for precise localization of
legged and non-legged locomotion, 2013
www.DLR.de • Chart 6 > IPIN 2013 > Patrick Robertson
Simultaneous Localization and Mapping (SLAM)
- Simultaneously Localizing and has been done in robotics since ~25 years - Smith, Self, Cheeseman (1990) - Leonard, Durrant-Whyte (1991)
- SLAM requires exteroceptive sensor to “see” the world- FootSLAM (2009) – Learns and uses a building layout while walking
around in it (odometry only)- PlaceSLAM (human labeling) (2010)- Robotic SLAM using the local magnetic field: Haverinen,Vallivaara et
al. (2009-2010)- ActionSLAM, Hardegger et al. (IPIN 2012, IPIN 2013)- Wifi-SLAM (Ferris, 2007) and WiSLAM (Bruno 2011, IPIN 2013)- SignalSLAM (Mirowski et al., IPIN 2013)
www.DLR.de • Chart 7 > IPIN 2013 > Patrick Robertson
Objectives for this Paper
- Magnetic SLAM for “free roaming” pedestrians in large
environments
- Foot mounted IMU / co-located magnetometer
- Bayesian Derivation
- Manageable complexity (real-time capability)
- Confirm it works in 3D
- Verification with accurate ground-truth to confirm low-decimeter
accuracy aspirations
www.DLR.de • Chart 8 > IPIN 2013 > Patrick Robertson
Basis for Derivation: Dynamic Bayesian Network
www.DLR.de • Chart 9 > IPIN 2013 > Patrick Robertson
- B: Magnetic environment map
- E: Error states of the odometry
- ZU: Measured odometry Step
- ZB: Measured Magnetic Field
- U: Actual step taken
(pose change vector)
- P: Pose
- Int, Vis: Human processing
- M: Physical environment map
“FootSLAM”
“MagSLAM”
MagSLAM Implementation with a Sequential Monte Carlo Filter
- FastSLAM factorization (Montemerlo et al. 2002)
- Proposal Function:
- propose odometry error process
- particles are propagated by odometry and their individual odometry
error
- Likelihood Functions:
- FootSLAM weighting (hexagonal edge crossing counters)
- and MagSLAM predictive posterior
- MagSLAM & FootSLAM per-particle mapping: Update each cell’s statistics
www.DLR.de • Chart 10 > IPIN 2013 > Patrick Robertson
Magnetic Field Strength at Ground Level: Finding a Robust Map Representation
www.DLR.de • Chart 11 > IPIN 2013 > Patrick Robertson
Source: M. Frassl et al., “Magnetic maps of indoor environments for precise localization of legged and non-legged locomotion,” in Intelligent Robots and Systems (IROS), Tokyo, Japan, Nov. 2013.
Hierarchical Magnetic Field Map Composed of Regular Grids
www.DLR.de • Chart 12 > IPIN 2013 > Patrick Robertson
Mapping: Spatial Binning of Measurements
www.DLR.de • Chart 13 > IPIN 2013 > Patrick Robertson
1: 65 mT
2: 63 mT
3: 94 mT
4: 74 mT
1: 74 mT
2: 84 mT
1: 84 mT
5: 84 mT
3: 87 mT
6: 87 mT 7: 42 mT
All we save are sample mean,
variance and N
Localization: Conjugate Bayesian Analysis
- Assume the measurements in each bin follow a normal distribution with
constant but unknown mean m and precision l- Assume a prior distribution on m and l that jointly follow the conjugate
prior of a normal distribution, i.e., a normal-gamma distribution:- NG(m, l | m0, k0, a0, b0)
- Then: Given the sufficient statistics of independent magnetic field
measurements taken within a bin, the posterior mean and precision also
jointly follow a normal-gamma distribution- The resulting posterior predictive of a new measurement is a student T-
distribution:
> IPIN 2013 > Patrick Robertsonwww.DLR.de • Chart 14
Evolution of the Posterior Predictive - No Measurements
www.DLR.de • Chart 15 > IPIN 2013 > Patrick Robertson
Prio
r P
DF
Evolution of the Posterior Predictive
www.DLR.de • Chart 16 > IPIN 2013 > Patrick Robertson
Pos
terio
r P
DF
His
togr
am p
er b
in
Evolution of the Posterior Predictive
www.DLR.de • Chart 17 > IPIN 2013 > Patrick Robertson
Pos
terio
r P
DF
His
togr
am p
er b
in
Results
www.DLR.de • Chart 18 > IPIN 2013 > Patrick Robertson
Video of Walking in Holodeck
www.DLR.de • Chart 19 > IPIN 2013 > Patrick Robertson
Comparison of MagSLAM Map with Ground Truth
www.DLR.de • Chart 20 > IPIN 2013 > Patrick Robertson
Not visited by
the human
Position Error Evolution
www.DLR.de • Chart 21 > IPIN 2013 > Patrick Robertson
Average error over 4 similar datasets: ~15 cm
MagSLAM Map of a Residential Flat
www.DLR.de • Chart 22 > IPIN 2013 > Patrick Robertson
Error at the Reference = Revisited Starting Location
www.DLR.de • Chart 23 > IPIN 2013 > Patrick Robertson
Error for MagSLAM alone: ~13 cm
Example in Large Office Buildings - 1
www.DLR.de • Chart 24 > IPIN 2013 > Patrick Robertson
met
ers
meters
MagSLAM from IPIN Demo
www.DLR.de • Chart 25 > IPIN 2013 > Patrick Robertson
met
ers
Example in Large Office Buildings - 2
www.DLR.de • Chart 26 > IPIN 2013 > Patrick Robertson
3D Video
Chart 27 > IPIN 2013 > Patrick Robertson
Thanks!Questions?
Videos on our website
www.kn-s.dlr.de/indoornav/magnetic.html
www.DLR.de • Chart 28 > IPIN 2013 > Patrick Robertson