closed-form algorithms in hybrid positioning: myths and misconceptions

16
1 Workshop for Positioning, Navigation and Communication 2010 11.06.22 Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions Niilo Sirola Department of Mathematics Tampere University of Technology, Finland (currently with Taipale Telematics, Finland) [email protected]

Upload: jody

Post on 25-Feb-2016

70 views

Category:

Documents


3 download

DESCRIPTION

Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions. Niilo Sirola Department of Mathematics Tampere University of Technology, Finland (currently with Taipale Telematics, Finland) [email protected]. Mobile positioning. Given a measurement model - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

1

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

Niilo SirolaDepartment of Mathematics

Tampere University of Technology, Finland

(currently with Taipale Telematics, Finland)[email protected]

Page 2: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

2

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Mobile positioning

Given a measurement model y = h(x) + vand the measurements y

Find the position x that fits the measurements ”best”

Page 3: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

3

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Iterative Least Squares

The Gauss-Newton or Taylor series or Iterative/Ordinary/Nonlinear least squares

Usual objections to Gauss-Newton1) initial guess: ”Selection of such a starting

point is not simple in practice”2) convergence is not assured3) computational load: ”as LS computation is

required in each iteration”

Page 4: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

4

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Examples of closed-form methods

geometrically inspired methods – easy to explain and visualise

”replace each intersection with a line

then solve the linear LS problem”

Page 5: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

5

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Examples of closed-form methods

others are algebraic and more rigorous

sometimes come with a proof

sometimes can be implemeted by the reader

Page 6: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

6

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Least-squares vs least-quartic solution

Least-squares solution:Find x such that ‖y – h(x)‖2 is as small as possible

Least-quartic solution:‖y2 – h(x) 2‖2

easier to solve analytically, but the solution is not least squares solution

-> is non-optimal in variance sense

Page 7: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

7

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Closed-form methods?

Some are not even in closed form….

- ”…first assume there is no relationship between x,y, and r1 … The final solution is obtained by imposing the relationship.. via another LS computation”

- ”we can first use (14) to obtain an initial solution…”

Page 8: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

8

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Testing method

-testing range-only methods-12 by 12 kilometer simulated test field, six ranging beacons-independent and identically distributed Gaussian measurement noise-noise sigma sweeps from 1 m to 10 km-1000 position fixes with random true position for each noise level

-5000 0 5000

-5000

0

5000

x / m

y / m

Page 9: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

9

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Tested algorithms

Candidate algorithms:1) ignore measurements – use the center of stations2) simple intersection3) range-Bancroft4) Gauss-Newton (with and without regularisation)5) Cheung (2006)6) Beck (2008)

All implemented in Matlab with similar level of optimization

Page 10: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

10

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Results: RMS position error

100

101

102

103

104

100

102

104

measurement noise / m

rmse / m

meansimplebancroftgauss-newtongauss-newton-regbeckcheung

Page 11: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

11

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Results: normalized error

100

101

102

103

104

0.5

1

1.5

2

measurement noise / m

norm

aliz

ed e

rror

meansimplebancroftgauss-newtongauss-newton-regbeckcheung

Page 12: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

12

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Back to the objections against GN…

1) requires an initial guess

so do several ”closed-form” methods

in practical applications rough position usually known from the context: physical constraints, station positions, etc.

possible to use a closed-form solution as a starting point

Page 13: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

13

Workshop for Positioning, Navigation and Communication 2010 22.04.23

2) Convergence not guaranteed

Depends on the quality of the initial guess

Regularisation helps

Sanity checks recommended - probably should use some with closed-form

methods as well!

Page 14: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

14

Workshop for Positioning, Navigation and Communication 2010 22.04.23

3) Computational complexity

Matlab on a 1.4GHz Celeron laptop

station mean: instantsimple intersection: 0.3 ms/fixBancroft: 0.5 ms/fix Cheung: 0.8 ms/fix Beck: 1.2 ms/fix

Gauss-Newton: 1.2 – 1.5 ms/fix

Page 15: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

15

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Bonus: flexibility

Closed-form methods:only ranges: okonly range differences: okmixed ranges and range differences: some choicesrange differences + a plane: at least one methodmixed ranges, range differences, planes, etc: … huh?

Gauss-Newton:combination of any (differentiable) measurements: OK

Page 16: Closed-form Algorithms in Hybrid Positioning: Myths and Misconceptions

16

Workshop for Positioning, Navigation and Communication 2010 22.04.23

Conclusions

Gauss-Newton was found to be competitive against several closed-form solutions

Additional bonus points:-Handles also correlated noise-Robust numerics-Gives an error estimates-Extends to time series -> Extended Kalman Filter

Questions?