LERTEKNIK

REG

Summary of Past Lectures

AU L
T O MA RO
T IC C O N T

Target Tracking: Lecture 4 Lecture-1

Multiple Target Tracking: Part I Introduction to and overview of target tracking.
Lecture-2
Emre Özkan Track initiation
[email protected] Single target data association
Division of Automatic Control Lecture-3
Department of Electrical Engineering How to detect maneuvers.
Linköping University
Linköping, Sweden Detection and multiple model based maneuver compensation.

November 25, 2014

E. Özkan Target Tracking November 25, 2014 1 / 29 E. Özkan Target Tracking November 25, 2014 2 / 29

Lecture Outline What is a hypothesis?

Definition
An (association) hypothesis is a partitioning of a set of measurements
according to the their origin.
What is an hypothesis?
At each time step, a single hypothesis tracking algorithm keeps only a
What is single hypothesis about all of the measurements received in the past.
• Single Hypothesis Tracking (SHT)?
• Global nearest neighbor algorithm does this by selecting the best
• Multiple Hypothesis Tracking (MHT)?
hypothesis according to a criterion.
Single Hypothesis Tracking • Joint probabilistic data association filter (JPDAF) combines all possible
• Global nearest neighbor (GNN) current hypotheses into a single one to form a single composite
• Joint probabilistic data association (JPDA) hypothesis. For this reason it can also be called as “composite
hypothesis tracking”.
A multiple hypothesis tracker, on the other hand, keeps multiple
hypotheses about the origin of the received data and has much more
computation and memory requirements.

E. Özkan Target Tracking November 25, 2014 3 / 29 E. Özkan Target Tracking November 25, 2014 4 / 29
Basic Scenario Considered in the Lecture Single hypothesis tracking

All the past is summarized by a 3 track hypothesis and possibly some

tentative tracks. All the past is summarized by a single hypothesis.
Tentative track processing is the same as what we learned in In this single hypothesis, we have nT tracks and nI initiators (or tentative
Lecture-2. tracks). Generally, the initiation procedure is separated from the main logic.
Using single target tracking methods for each target gives only locally When a set of new measurements arrives, one first gates the measurements with
optimal results. the existing (confirmed) targets.
Using the gating results, Yk
association is carried out.
The global picture must be yk2 List List
yk1 3 Using association results, of Gating Association Update of
ŷk|k−1
taken into account for targets confirmed tracks are updated.
Tracks Tracks

ed
m
Unprocessed
Yk

ir
Measurements

sharing measurements in their

f
on
C
If
Unprocessed remaining List
Initiator
List
gates or possibly some other 1
ŷk|k−1
of
Initiators
Processing
of
Initiators
yk3 2
ŷk|k−1 measurements are sent to the
measurement-to-target k k+1
time
initiator logic.
association conflict. yk4

E. Özkan Target Tracking November 25, 2014 5 / 29 E. Özkan Target Tracking November 25, 2014 6 / 29

Gating Association Hypotheses

Iterate over measurements
yk1 yk2 yk3 yk4
Suppose there are nT = 3 (confirmed) tracks in the hypothesis measurements
summarizing the past. Once we get the measurements T1
T2
FA yk1
yk2
3
ŷk|k−1
Yk = {yk1 , . . . , yk4 }, using the gate criteria we can prepare the NT

following matrix to facilitate hypothesis generation. T2 FA

NT
FA 1
ŷk|k−1
T2 2
FA FA yk3 ŷk|k−1
NT
T2 yk4

Hypotheses
NT FA
T1 T2 T3 yk2 NT
yk1 3 T3
ŷk|k−1 T2
1 0 0 1 T1 FA Validation Matrix
NT
2 0 0 0 T1 T2 T3
T2 FA
3 1 1 0 NT 1 0 0 1
1 NT
4 0 1 0 ŷk|k−1
2
T2 2 0 0 0
yk3 ŷk|k−1 FA FA
Such a matrix is called as NT
3 1 1 0
T2
validation matrix. yk4 NT FA 4 0 1 0
NT

Repeat the procedure above for yk1 = FA and yk1 = NT.

E. Özkan Target Tracking November 25, 2014 7 / 29 E. Özkan Target Tracking November 25, 2014 8 / 29
 Association Hypotheses cont’d Probability of a Hypothesis

We can define an association hypothesis θk formally as a mapping Suppose we are at time k at an intermediate stage of tracking. We have
j = 1, . . . , nT targets established previously and have just received
θk (·) : {1, 2, . . . , mk } → {FA, 1, 2, . . . , nT , NT} Yk = {yk1 , . . . , ykmk }
where mk is the number of measurements in Yk i.e., Yk = {yk1 , . . . , ykmk } Suppose θk (·) is an arbitrary hypothesis about the origin of Yk .
nT is the number of targets formed in the past. Number of false alarms mFk A in S, the surveillance region is
distributed as PF A (mFk A );
Example Hypotheses with mk = 4, nT = 3
False alarm spatial density is pF A (y)
FA FA FA yk1
yk2 Number of new targets in S, the surveillance region is distributed as
1 1 1 3
ŷk|k−1
PN T (mN T
1 1 1 k );
2 2 2 New target spatial density is pN T (y);
2 2 2
Detection probability of the jth target: PDj ;
1
ŷk|k−1
3 3 3 yk3 2
ŷk|k−1
3 3 3 yk4 Gate probability of the jth target: PGj ;
4 4 4
NT NT NT Predicted measurement density of jth target: pjk|k−1 (y).

E. Özkan Target Tracking November 25, 2014 9 / 29 E. Özkan Target Tracking November 25, 2014 10 / 29

Standard Settings cont’d Fundamental Theorem of TT

Theorem: Suppose θk is an association hypothesis about the current measurement

set Yk . Then the posterior probability of θk is given as
  
FA
(βF A VS )mk exp(βF A VS )
 −1 
F A N T
PF A (mFk A ) = m m θ (j)
PDj pjk|k−1 ykk (1 − PDj PGj )
Y Y
mF A
k !
P (θk |Y0:k ) ∝ βF Ak βN Tk  
pF A (y) = 1/VS when y ∈ VS . j∈JD j∈JN D
NT
(βN T VS )mk exp(βN T VS ) where
PN T (mN T
k )= mN T
k !
JD is the set of indices of detected targets, i.e., indices of targets which were
pN T (y) = 1/VS when y ∈ VS .
assigned a measurement by θk .
JN D is the set of indices of non-detected targets i.e., indices of target that were
not assigned a measurement by θk .
θk−1 (j) is the index of the measurements that is assigned to target when j ∈ JD .

E. Özkan Target Tracking November 25, 2014 11 / 29 E. Özkan Target Tracking November 25, 2014 12 / 29
 Fundamental Theorem of TT cont’d Fundamental Theorem of TT cont’d

Since there is a single hypothesis for the past, the term We can write the summed elements in the log-probability in a matrix given as
nT
(1 − PDj PGj ) = (1 − PDj PGj ) (1 − PDj PGj )
Y Y Y
A T1 T2 T3 FA1 FA2 FA3 FA4 NT1 NT2 NT3 NT4
j=1 j∈JD j∈JN D
yk1 × × `13 log βF A × × × log βN T × × ×
is constant for all hypotheses. Then, we have yk2 × × × × log βF A × × × log βN T × ×
 −1 
θk (j)
yk3 `31 `32 × × × log βF A × × × log βN T ×
j j
mF A mN T Y PD pk|k−1 yk yk4 × `42 × × × × log βF A × × × log βN T

P (θk |Y0:k ) ∝ βF Ak βN Tk where

j∈JD (1 − PDj PGj )
× represents −∞. Finding the optimal association hypothesis then
P j pj (yki ) corresponds to finding the column indices
Taking the logarithm, we have  −1  `ij , log D k|k−1 .
j j
θ (j) {j1 , j2 , j3 , j4 }, ji1 6=
Pj4i2 for 1 ≤ i1 6= i2 ≤ 4
j j (1−PD PG )
PD pk|k−1 ykk
X This matrix is called such that the sum i=1 Aiji is maximized.
log P (θk |Y0:k ) = mFk A log βF A+mN T
k log βN T + log
j∈JD (1 − PDj PGj ) assignment matrix.

E. Özkan Target Tracking November 25, 2014 13 / 29 E. Özkan Target Tracking November 25, 2014 14 / 29

Assignment Problem Assignment Problem cont’d

We can make a formal definition of the problem as follows

We are given the matrix A ∈ Rm×n with n ≥ m. Considered first in economics.
Define the auxiliary matrix Z = [zij ] where zij ∈ {0, 1}. Possible equivalents are assigning personnel to jobs or assigning
delivery trucks to locations.
Problem Definition
Earlier methods used linear programming techniques, like Hungarian
m
X method which is computationally costly.
Maximize zij Aij
i=1
Less computationally expensive methods appeared later when
subject to different applications were found.
n
X X m • Munkres algorithm
zij = 1 ∀i and zij ≤ 1 ∀j • JVC algorithm (by Jonker and Volgenant)
j=1 i=1 • Auction algorithm (by Bertsekas): Explained thoroughly in the book.

This problem is called as assignment problem in optimization literature.

E. Özkan Target Tracking November 25, 2014 15 / 29 E. Özkan Target Tracking November 25, 2014 16 / 29
Global Nearest Neighbor (GNN) Algorithm GNN Algorithm cont’d

Therefore, we combine the FA and NT cases into the single category of external
sources (EX).
A T1 T2 T3 FA1 FA2 FA3 FA4 NT1 NT2 NT3 NT4 The external sources become Poisson with density βEX = βF A + βN T .
yk1 × × `13 log βF A × × × log βN T × × ×
yk2 × × × × log βF A × × × log βN T × × A T1 T2 T3 FA1 FA2 FA3 FA4 NT1 NT2 NT3 NT4
yk3 `31 `32 × × × log βF A × × × log βN T × yk1 × × `13 log βF A × × × log βN T × × ×
yk4 × `42 × × × × log βF A × × × log βN T
yk2 × × × × log βF A × × × log βN T × ×
yk3 `31 `32 × × × log βF A × × × log βN T ×
Choose the best (largest probability) association hypothesis. yk4 × `42 × × × × log βF A × × × log βN T

The measurements associated to targets in the best association hypothesis are A T1 T2 T3 EX1 EX2 EX3 EX4
processed by target KFs. yk1 × × `13 log βEX × × ×
The measurements associated to FA and NT are handled by the initiator logic. yk2 × × × × log βEX × ×
yk3 `31 `32 × × × log βEX ×
yk4 × `42 × × × × log βEX

E. Özkan Target Tracking November 25, 2014 17 / 29 E. Özkan Target Tracking November 25, 2014 18 / 29

Global Nearest Neighbor (GNN) Algorithm PDA vs. JPDA

GNN
Time k = 0: Send all measurements to initiation logic.
Time k > 0: Suppose we have mk measurements and nT targets
• Form the assignment matrix A ∈ Rmk ×(nT +mk )
• auction(A)
• Process the targets with their associated measurements.
• Send the measurements associated to external sources (EX) to
initiation logic.
• Process the initiators (tentative tracks with EX associated
measurements).
• Add any confirmed tentative track to the confirmed track list.

Figure taken from:

R.J. Fitzgerald, “Development of Practical PDA Logic for Multitarget Tracking by Microprocessor,”
American Control Conference, pp.889–898, Jun. 1986.

E. Özkan Target Tracking November 25, 2014 19 / 29 E. Özkan Target Tracking November 25, 2014 20 / 29
 Joint Probabilistic Data Association (JPDA) Filter JPDAF cont’d

Soft decision equivalent of GNN in the way that PDA is a soft version yk1 yk2 yk3 yk4
of NN. measurements
T2 p1 ∝ βF A `13 `31 `42
All past is again summarized with a single hypothesis (nT confirmed T1
FA p2 ∝ βF2 A `13 `31 `¯2
targets nI tentative targets).
T3 FA T2 FA p3 ∝ βF2 A `13 `32 `¯1
The number of targets is assumed fixed in the association, hence no
FA T2 p4 ∝ βF2 A `13 `42 `¯1 yk2
N T associations in the possible hypotheses.

Hypotheses
yk1 3
ŷk|k−1
FA p5 ∝ βF3 A `13 `¯1 `¯2
For each previously established target, we need to calculate
T2 p6 ∝ βF2 A `31 `42 `¯3 1
ŷk|k−1
T1 yk3 2
ŷk|k−1
P (Tj ↔ yki ) and P (Tj ↔ φ) (1) FA p7 ∝ βF3 A `31 `¯2 `¯3
yk4
FA FA T2 FA p8 ∝ βF3 A `32 `¯1 `¯3
for yki that are in the gate of the target. The update is then made FA T2 p9 ∝ βF3 A `42 `¯1 `¯3
with PDA update formulas by using these probabilities instead. p10 ∝ βF4 A `¯1 `¯2 `¯3
FA
A separate track initiation logic must run along with JPDAF to detect
and initiate new tracks. `ij = PDj pjk|k−1(yki ) `¯j = 1 − PDj PGj

E. Özkan Target Tracking November 25, 2014 21 / 29 E. Özkan Target Tracking November 25, 2014 22 / 29

JPDAF cont’d JPDAF cont’d

T1 T2 T3
measurements Probability calculations show that the calculations can be done
P (T1 ↔yk3 )= 1c
P separately for the clusters of targets that does not share gated
yk1 p1 ∝ βF A `13 `31 `42 i∈{1,6,2,7} pi
yk4 measurements.
φ p6 ∝ βF2 A `31 `42 `¯3 P (T1 ↔φ)= 1c
P
i∈{3,4,5,8,9,10} pi
In other words our previous scenario can be grouped into two clusters
yk3 yk1 p2 ∝ βF2 A `13 `31 `¯2
φ T1 &T2 , T3 and probability calculations can be done separately for the
φ p7 ∝ βF3 A `31 `¯2 `¯3 P (T2 ↔yk3 )= 1c
P
pi corresponding hypothesis trees.
Hypotheses

i∈{3,8}

yk1 p3 ∝ βF2 A `13 `32 `¯1 P (T2 ↔yk4 )= 1c

P
i∈{1,6,4,9} pi
yk3
φ p8 ∝ βF3 A `32 `¯1 `¯3 P (T2 ↔φ)= 1c
P
i∈{2,7,5,10} pi

yk1 p4 ∝ βF2 A `13 `42 `¯1 yk1

yk2 The clustering can be done using the
φ yk4 3
ŷk|k−1 T1 T2 T3
φ p9 ∝ βF3 A `42 `¯1 `¯3 P (T3 ↔yk1 )= 1c
P5
pi
i=1
1 0 0 1
φ yk1 p5 ∝ βF3 A `13 `¯1 `¯2
P10
P (T3 ↔φ)= 1c i=6 pi validation matrix 2 0 0 0
1
ŷk|k−1
φ p10 ∝ βF4 A `¯1 `¯2 `¯3 yk3 2
ŷk|k−1 3 1 1 0
c=
P10
i=1 pi yk4
4 0 1 0
`ij = PDj pjk|k−1(yki ) `¯j = 1 − PDj PGj

E. Özkan Target Tracking November 25, 2014 23 / 29 E. Özkan Target Tracking November 25, 2014 24 / 29
 JPDAF cont’d JPDAF cont’d
T1 T2 T3
P (T1 ↔yk3 )= c1 (p1 +p2 )
measurements 1
JPDAF
yk1 P (T1 ↔φ)= c1 (p3 +p4 +p5 )
p6 ∝ `13 1
Time k = 0: Send all measurements to initiation logic.
φ p7 ∝ βF A `¯3 Time k > 0: Suppose we have mk measurements and nT targets
P (T2 ↔yk3 )= c1 p3 • Form the validation matrix.
yk4 p1 ∝ `31`42 1
yk3 • Group the targets into clusters in which targets share gated
P (T2 ↔yk4 )= c1 (p1 +p4 )
Hypotheses

1
φ p2 ∝ βF A `31 `¯2 measurements.
P (T2 ↔φ)= c1 (p2 +p5 ) • For each cluster, calculate PDA probabilities for each target in the
1
yk3 p3 ∝ βF A `32 `¯1 cluster by using a hypothesis tree.
φ • Update targets with the weighted equivalent measurements as PDA.
yk4 p4 ∝ βF A `42 `¯1 P (T3 ↔yk1 )= c1 p6
2 • Send the unprocessed measurements and possibly gated extra
φ p5 ∝ βF2 A `¯1 `¯2 P (T3 ↔φ)= c1 p7 measurements to initiation logic.
2
P5 • Process the initiators.
c1 = pi • Add any confirmed tentative track to the confirmed track list.
`ij = PDj pjk|k−1(yki ) `¯j = 1 − PDj PGj i=1

c2 =p6 +p7

E. Özkan Target Tracking November 25, 2014 25 / 29 E. Özkan Target Tracking November 25, 2014 26 / 29

Track Coalescence Problem of JPDAF References

R. J. Fitzgerald, “Development of Practical PDA Logic for Multitarget Tracking by

Microprocessor,” American Control Conference, pp.889–898, Jun. 1986.
H. A. P. Blom, and E. A. Bloem, “Probabilistic data association avoiding track
coalescence,” IEEE Transactions on Automatic Control, vol.45, no.2, pp.247–259, Feb.
2000.
I. J. Cox and S. L. Hingorani, “An efficient implementation of Reid’s multiple hypothesis
tracking algorithm and its evaluation for the purpose of visual tracking” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol.18, no.2, pp.138–150,
Feb. 1996.
R. Danchick and G. E. Newnam, “Reformulating Reid’s MHT method with generalised
Murty K-best ranked linear assignment algorithm,” IEE Proceedings – Radar, Sonar and
Navigation, vol.153, no.1, pp. 13–22, 16 Feb. 2006.
Figures taken from:
R.J. Fitzgerald, “Development of Practical PDA Logic for Multitarget Tracking by Microprocessor,”
American Control Conference, pp.889–898, Jun. 1986.

E. Özkan Target Tracking November 25, 2014 27 / 29 E. Özkan Target Tracking November 25, 2014 28 / 29
References

S. Blackman and R. Popoli, Design and Analysis of Modern Tracking Systems.

Norwood, MA: Artech House, 1999.
Y. Bar-Shalom and X. R. Li, Multitarget-Multisensor Tracking: Principles, Techniques.
Storrs, CT: YBS Publishing, 1995.

E. Özkan Target Tracking November 25, 2014 29 / 29