This document presents a method for determining angular probability density functions (PDFs) to be used as a transition model in a particle filter for indoor navigation. It calculates diffusion matrices based on a user's path through a building to determine likely step directions. The angular PDF is directly determined from the diffusion matrix and favors angles without wall crossings. Simulation results show the angular PDF approach improves accuracy over using just wall constraints, particularly in multi-modal scenarios where particles could diverge. The method provides a non-binary way to incorporate map information into the particle filter weighting.
1. A Maps-Based Angular PDF for Navigation Systems
in Indoor and Outdoor Environments
Susanna Kaiser, Mohammed Khider, Patrick Robertson
German Aerospace Center (DLR)
Institute of Communications and Navigation
Center of Excellence for Satellite Navigation
Folie 1
Vortrag > Kaiser > IPIN2011.ppt > 09.11.2005
2. Outline
Indoor Positioning System
Motivation for Using Angular PDFs
Algorithm for Determining Angular PDFs
Calculation of the Diffusion Matrix
Determining the angular PDF directly from the diffusion matrix
Simulation Results
Conclusions & Outlook
2
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
3. Indoor Positioning System
Cascaded Bayesian Estimation Architecture:
GNSS
Pseudoranges Likelihood Particles,
Functions: Sensor errors Position and Velocity Output
& Carriers
GNSS, Fusion
Altimeter Outputs Altimeter, Filter Particles
RFID Detections RFID (PF)
Compass, Transition
Model 3D Map
Compass Outputs Step/INS
~1 Hz Database
Fusion Trigger
Step Displacement INS Position and
( P) Velocity PDF
P PDF (Gaussian) Computer
Gyroscope INS Position and Velocity
Triad Outputs Strapdown +
Accelerometer Inertial Calibration Feedback -
Triad Outputs Computer
~100-500 Hz INS Errors PDF
Extended
Sensors “Foot still” – Detector
Kalman Filter
Triggers ZUPT
(INS Error Space)
3
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
4. Bayesian Filter Algorithm
Two stages:
Prediction, transition model is used to predict the state PDF
from one time-step to the next
p(xk | z1:k 1) p(xk | xk 1) p(xk 1 | z1:k 1)dxk 1
Update, measurement model is used to update the prediction PDF
according to the latest measurement
p ( xk | z1:k ) k p ( z k | xk ) p ( xk | z1:k 1 )
Several types of Bayesian filters results in different accuracy, robustness,
efficiency, sensor variety and complexity
Likelihood Particle Filter: uses measurements for importance sampling and
transition model to update (for weighting)
4
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
5. Motivation for Using Angular PDFs
Likelihood Particle Filter is used -> We need the Transition Model for
weighting the Particles. (But: The angular PDF can also be used as a
movement model!)
The main problem of the whole system is the heading drift, therefore,
we concentrate our focus on the heading.
The heading of a walking pedestrian is correlated with maps or floor-
plans.
Floor-plans are used in a binary way: if a particle crosses a wall it‘s
weight is set to zero.
This may lead to estimation failures in multi-modal scenarios
(e.g. the “Y-problem”)
How can we find a weighting function that uses floor-plans in a non
binary way? A transition model will not just eliminate particles that
cross walls but rather reward those that follow a trajectory
compatible with the building layout.
Advantages to the previous used Diffusion Movement Model:
No destination points, no path finding algorithm, smaller area
5
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
6. Motivation for using angular PDFs: The „Y-Problem“:
Building Building
Part A Part B
True track
Particles
Split
Point
Particles in A will be eliminated after time due to the wall constraints
6
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
7. Motivation for using angular PDFs:
Example for a Critical Multi-Modal Scenario Based on
Our Own Office Environment
x ground truth point
path of the test user
x6
Location where the x7
particle cloud splits x5
into two groups x8
x4
x9
x3
Large room
x2
x1 Start/End
Particles will be split into 2 groups at GTP4.
Particles that enter the large room will dominate due to the wall constraints when
using binary weighting.
7
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
8. Angular PDFs:
Calculation of the Diffusion Matrix
For each actual waypoint a sliding squared window of size Nx Nx is defined,
where the waypoint is the middle point of that window.
Each waypoint represents a source effusing gas.
Source/waypoint
Example for a diffusion matrix
For each waypoint x m , ym a diffusion matrix D is pre-computed.
x
At start, the diffusion matrix is one at the waypoint and zero else where.
8
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
9. Angular PDFs:
Determining the angular PDF directly out of the diffusion matrix
From the diffusion matrix a threshold can be used for obtaining a contour line of the
gas distribution.
Source/waypoint
Contour line (dark red) of the diffusion values with threshold value: T=0.0001
Resulting in a set C of Nc contour-line points:
c1,..., cN
x1, y1 ,, xN , yN
c
c c
Contour Points (dark red) are included in the set if
The direct line from the source to the contour point is not crossing a wall.
At least one neighboring value is greater than the threshold.
9
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
10. Angular PDFs:
Determining the angular PDF directly out of the diffusion matrix
For each angle the distance from the middle waypoint to the contour point is determined
and the maximum distance is used:
0°
b (xm k )2 ( ym l)2
C(k,l)
Source/waypoint
b
w( ) max bC (k ,l )
C (k ,l )
(k ,l )
Normalizing: w( ) w( )
2
w( )
0
Adapting to the speed S: w'( ) w( )S
10
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
11. Angular PDFs:
Determining the angular PDF directly out of the diffusion matrix
Resulting Polar Plot of the angular PDF for that specific waypoint:
By using this weighting function within the LPF angles in the possible
direction without wall constraints are favored.
11
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
12. Simulation Results
Simulation Environment
Sequential Bayesian Positioning Estimator:
Based on a Particle Filter fusion engine.
Integrates the new angular PDFs.
Using the following sensors:
Commercial GPS
Electronic compass
A foot-mounted Inertial Measurement Unit (IMU) with Zero
Velocity Updates (ZUPTs) processed with an extended Kalman
Filter for pedestrian dead reckoning [Krach and Robertson].
The ground truth points (GTRP) were measured to the sub-
centimeter accuracy using a tachymeter (Leica Smart Station (TPS
1200), employing optical distance, angular measurements, differential
GPS (initial positioning)).
100 runs for a single walk (playback mode)
12
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
13. Simulation Results
Average Position Error (Errors between the true and estimated pedestrian
positions)
Average Position Error with wall constraints: 1.33m (no angular PDF:
1.5m)
13
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
14. Simulation Results
Number of Failures
Rates of wrong position estimations of the scenario
Location of wrong position estimation
Algorithm
Wrong position in first loop Wrong position in
second/third loop
Angular PDF 0% 8%
No angular PDF 100% 50%
14
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
15. Conclusions & Outlook
A transition model for pedestrians that uses a known building layout to
construct an angular PDF for a pedestrian’s step direction.
A simple Likelihood Particle Filter (using binary weighting) can diverge if
the particles distribution is multi-modal and a competing, erroneous
particles group is in an area with few limiting walls in their vicinity.
It has been shown that weighting with the angular PDF performs better
than using no transition model in critical scenarios.
Further work should focus on
More data sets
Different wall situations
The use of different accessibility levels in indoor and outdoor
environments.
15
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011
16. Many thanks for your interest
&
your questions are welcome!
http://www.kn-s.dlr.de/indoornav
Susanna Kaiser
Date: 22/09/2011
16
A Maps-Based Angular PDF … > Kaiser > IPIN2011 > 22.09.2011