Sparse Kalman Filter

Joan Solà

April 23, 2010

1 Prediction-type operations
1.1 Robot motion
General function x = f (x, u, q)
Sparse function r = f (r, u, q)
In r robot
Invariant m mapped landmarks
Out r robot
Ignored in i
Used in x=r+m
Used out x=r+m
Ignored out i

Matrix partitions  
x= r m i
Jacobian  
Fr 0 ∗
F =  0 I ∗
∗ ∗ ∗
Covariances  
Prr Prm ∗
P = Pmr Pmm ∗
∗ ∗ ∗
Output covariances

Fr Prr Fr> + Qrr Fr Prm ∗

FPF> + Q =  Pmr Fr> Pmm ∗

1
1.2 Landmark initialization
General function x = g(x, y, n)
Sparse function l = g(r, y, n)
In r robot+sensor
Invariant r+m robot+sensor+map
Out l new landmark
Ignored in i
Used in x=r+m
Used out x=r+m+l
Ignored out i−l

Matrix partitions  
x= r m l i
Jacobian  
I 0 ∗ ∗
0 I ∗ ∗
G= 
Gr 0 ∗ ∗
∗ ∗ ∗ ∗
Covariances  
Prr Prm ∗ ∗
Pmr Pmm ∗ ∗
P =
 ∗

∗ ∗ ∗
∗ ∗ ∗ ∗
Output covariances

Prr G>
 
Prr Prm r ∗
P P P G > ∗
mr mm mr
GP G> + (R) + (N ) = 
 r
Gr Prr Gr Prm Gr Prr Gr + Gy RG>
> >

y + Gn N Gn ∗
∗ ∗ ∗ ∗

2
1.3 Landmark re-parametrization
General function x = j(x)
Sparse function l = j(k)
In k old landmark
Invariant m all map
Out l new landmark
Ignored in i
Used in x=k+m map with old lmk
Used out x=m+l map with new lmk
Ignored out i+k−l

Matrix partitions  
x= k m l i
Jacobian  
∗ ∗ ∗ ∗
0 I ∗ ∗
J = 
Jk 0 ∗ ∗
∗ ∗ ∗ ∗
Covariances  
Pkk Pkm ∗ ∗
Pmk Pmm ∗ ∗
P =
 ∗

∗ ∗ ∗
∗ ∗ ∗ ∗
Output covariances  
∗ ∗ ∗ ∗
∗ Pmm P mk Jk> ∗
JP J > =
∗ Jk Pkm Jk Pkk J >

k ∗
∗ ∗ ∗ ∗

3
2 Correction-type operations
2.1 Individual landmark correction
General function y = h(x)
Sparse function y = h(r, l)
In 1 r robot+sensor
In 2 l observed landmark
Passive m other landmarks
Out y measurement
Updated r+m+l robot+sensor+landmarks
Ignored i

Matrix partitions  
x= k m l i
Jacobian  
H = Hr 0 Hl ∗
Covariances  
Prr Prm Prl ∗
Pmr Pmm Pml ∗
P = 
 Plr Plm Pll ∗
∗ ∗ ∗ ∗
Expectation matrix

E = Hr Prr Hr> + Hr Prl Hl> + Hl Plr Hr> + Hl Pll Hl>

Innovation matrix
Z =E+R
Band matrix
Pxr Hr> + Pxl Hl>
 
P H> =
∗
Kalman gain
K = P H > Z −1
Covariances update
P = P − K(P H > )>

4
2.2 Buffered landmarks correction
General function y = h(x)
Sparse function y = h(r, l1 , l2 )
In 1 r robot+sensor
In 2 l1 observed landmark
In 3 l2 observed landmark
Passive m other landmarks
Out y1 measurement
Out y2 measurement
Updated r + m + l1 + l2 robot + sensor + all landmarks
Ignored i

Matrix partitions  
x = r m l1 l2 i
Jacobian  
Hr1 0 Hl1 0 ∗
H=
Hr2 0 0 Hl2 ∗
Covariances  
Prr Prm Pr1 Pr2 ∗
Pmr Pmm Pm1 Pm2
 ∗

 P1r P1m P11 P12
P = ∗

 P2r P2m P21 P22 ∗
∗ ∗ ∗ ∗ ∗
Expectation matrix  
E11 E12
E=
E21 E22
with
>
Eij = Hri Prr Hrj >
+ Hri Prj Hlj> + Hli Pij Hlj> + Hli Pir Hrj
Innovation matrix  
E11 + R E12
Z=
E21 E22 + R
Band matrix
P H> (P H > )1 (P H > )2
   
=
∗ ∗ ∗
with
(P H > )i = Pxr Hri
> + P H>
 
xi li

Kalman gain
P H > −1
   
K
= Z
∗ ∗
Covariances update
P = P − K(P H > )>