Improving Simultaneous Localization and Mapping for Pedestrian Navigation and Automatic Mapping of Buildings by using Online Human-Based Feature Labeling
Improving Simultaneous Localization and Mapping for Pedestrian Navigation and Automatic Mapping of Buildings by using Online Human-Based Feature Labeling.
Pesented at IEEE/ION PLANS 2010, May 2010, Palm Springs, CA, USA.
Authors: Patrick Robertson, Michael Angermann, Mohammed Khider, German Aerospace Center (DLR)
Similar to Improving Simultaneous Localization and Mapping for Pedestrian Navigation and Automatic Mapping of Buildings by using Online Human-Based Feature Labeling
Similar to Improving Simultaneous Localization and Mapping for Pedestrian Navigation and Automatic Mapping of Buildings by using Online Human-Based Feature Labeling (16)
The Future of Software Development - Devin AI Innovative Approach.pdf
Improving Simultaneous Localization and Mapping for Pedestrian Navigation and Automatic Mapping of Buildings by using Online Human-Based Feature Labeling
1. Improving Simultaneous Localization and Mapping for
Pedestrian Navigation and Automatic Mapping of Buildings
by using Online Human-Based Feature Labeling
Patrick Robertson, Michael Angermann,
Mohammed Khider, German Aerospace Center (DLR)
Slides from: “Improving Simultaneous Localization and Mapping for Pedestrian Navigation
and Automatic Mapping of Buildings by using Online Human-Based Feature Labeling”, in
Proc. IEEE/ION PLANS 2010, May 2010, Palm Springs, CA, USA.
2. SLAM in Robotics
Simultaneous Localization and Mapping - identified by
robotics community in mid ‘80s!
Premise:
Localization using odometry and sensing of known
landmarks is easy!
Mapping of landmarks given known location and
orientation (pose) is easy!
Simultaneous Localization and Mapping is hard!
3. What about SLAM for Humans?
Human pedestrians are not robots but share
some similarities with them
Visual sensors (eyes)
'Odometry' (in humans: sensed by
proprioception), can be measured
using inertial sensors
Path and planning and execution
For humans: little or no direct 'access' to
senses and functions
Our central assumption:
The pedestrian is able to actively control
motion without violating physical
constraints (i.e. walls, etc)
4. Raw NavShoe Odometry Results
NavShoe INS produced reasonable results
stand alone, but still unbounded error growth
NavShoe INS had larger heading slips;
unbounded error begins to rise earlier
Algorithm: Extended Kalman Filter with Zero Velcocity Updates (Foxlin)
6. Six ways out of the hexagon
First order Markov process
Location dependent
Time Invariant
Probabilistic map
FootSLAM: Hexagonal Grid over Space
Human motion is modelled by a person choosing
which edge of the hexagon to cross.
9. A
C
G
E
Physical space
F
B
D
1
2
3 4
5
6
7
8 9
10
11
A B C A F E B D G E D B
Timestamped placestamps
Perfect association
Partial association
Unknown association
- Arrows denote pedestian‘s trajectory;
- letter-coded circles with denote unique places;
- colors denote some recognizable aspect of the place
An Example of Placestamps
10. The PlaceSLAM Dynamic Bayesian Network (DBN)
P
U
Zu
E
Int
Vis
L, M
“Visual impression -
what the person sees“
Intention
“where the person
wants to go”
Time k-1
Measured
Step
“Environment” = Human recognizable
Places L and FootSLAM Map M; both are
constant over time
Time k
P
U
Zu
E
Int
Vis
ZL
A
Placestamp
Place
identifier
seen
ZL
A
Odometry
Error
states
Actual step taken
(pose change vector)
Pose (= location, orientation)
11. Intuitive Explanation of the Sequential
Monte Carlo Estimator
FootSLAM lets particles, or hypotheses, explore the state
space of odometry errors, like evolution of drift as well as
the association of places
In this way, every particle is trying a slightly “differently bent
piece of wire”
Particles are weighted by their “compatibility” with
their individual PlaceSLAM map
their individual FootSLAM map
optional sensor readings, such as GPS,
magnetometer
We can show that this is optimal in the Bayesian sense!
12. Illustration of Proposal Function 1
dmin
Particle position
dmin
If particle is closer than dmin to some existing place(s) then
choose the closest place
13. Illustration of Proposal Function 2
dmin
If particle is further than dmin from all existing places then
choose a new place at the particle‘s current position
Particle position
New place proposed to be here
14. Algorithm Summary
Perform
FootSLAM
Weighting
and FootSLAM
map update
Locate closest
existing place
to particle’s
current Pose P
Placestamp
was reported
Select this
Identifier
(closest place)
Choose
new identifier
None within
dmin
Closest is
within dmin
Multiply weight
by PL
Multiply weight
by Gaussian
Likelihood
(PLANS paper (12))
Initialise new place’s
location to current
particle pose P
Update place’s
location with current
particle pose P
Noplacestamp
reported
Perform
for all Np
Particles:
15. Weight update
If particle i revisited a place:
If particle marked a new place:
r cancels out and pL accounts for places being sparse
19. Experiments and Results
Measurement data taken from a pedestrian wearing a foot
mounted IMU
Placestamps collected during the walk
Two scenarios:
Indoor only
Outdoor – indoor - outdoor sequence
Indoor only: only foot mounted IMU
Mixed scenario: foot mounted IMU as well as GPS and
compass sensors
23. Concluding Notes
PlaceSLAM is a useful adjunct to FootSLAM and improves accuracy and stability
Two main forms of PlaceSLAM: Perfect Association (“press a certain button”) and
unknown association (“press any button”)
Error assumptions: Humans are lazy in reporting but do not erroneously report places
Bayesian derivation
Suggested future work:
More experimental data in different sites and for different building sizes and
geometries
Map building with multiple users; “crowdsourcing” collaborative mapping
Extend error models, overlapping and multiple places, RFID tags