Professional Documents
Culture Documents
Lecture 5
Lecture 5
REG
Lecture Outline
AU L
T O MA RO
T IC C O N T
1. Conceptual MHT
• Fundamental Components
• Simplifications
Target Tracking: Lecture 5 • Summary
Multiple Target Tracking: Part II 2. Hypothesis-Based MHT
• Assignment Problem
• Algorithm
3. Track-Based MHT
Gustaf Hendeby • Implementation Details
hendeby@isy.liu.se • Summary
4. User Interaction
Div. Automatic Control
Dept. Electrical Engineering 5. Examples
Linköping University • SUPPORT
• ADABTS
December 10, 2014 6. Summary
• Concluding Remarks
• Learn More. . .
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 1 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 2 / 36
Intro multi-target tracking (MTT) MHT: consider multiple associations hypotheses over time
Single hypothesis tracker (SHT) Started with the conceptual MHT
• Global nearest neighbor (GNN)
Integrated track initialization
• Joint Probabilistic Data Association (JPDA)
Two principal implementations
Auction algorithm • hypotheses based
Fundamental theorem of target tracking • track based
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 3 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 4 / 36
Conceptual MHT Conceptual MHT Fundamental Components
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 5 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 6 / 36
Hypotheses
FA FA
yk1 yk1 T1 NT T1 NT
2
ŷk|k−1 FA FA
1 1
ŷk|k−1 ŷk|k−1 FA NT FA NT
FA FA
yk2 yk2 T2 NT NT T1 NT NT
FA
T1 NT
FA
FA NT
FA
T1 NT NT
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 6 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 7 / 36
Conceptual MHT Fundamental Components Conceptual MHT Fundamental Components
New Set of
Measurements
Let Θ`k , {θk , Θik−1 }, then {yki }m k
i=1
(using the “Fundamental Theorem of TT”)
Reduce
P(Θ`k |y0:k ) ∝ p(yk |Θ`k , y0:k−1 )P(θk |Θik−1 , y0:k−1 )P(Θik−1 |y0:k−1 ) Set of
Hypotheses
Generate New
Hypotheses
Calculate
Hyp. Probabilities Number of
{Θik−1 }N {Θik }N {P (Θik )}N Hypotheses Θik
Y h h h
mkfa mknt Y j j θk−1 (j) j j i=1 i=1 i=1
∝ βfa βnt PD pk|k−1 (yk ) (1 − PD PG ) P(Θik−1 |y0:k−1 )
j∈JDi i
j∈JND
z −1
Note User
Presentation
The sets JDi and JNDi depend on Θik−1 ! The number of targets and target Logic
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 8 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 9 / 36
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 10 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 11 / 36
Conceptual MHT Simplifications Conceptual MHT Simplifications
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 11 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 12 / 36
Clustering: process clusters separately (1/2) Clustering: process clusters separately (1/2)
Hypotheses generation Hypotheses generation
Form Θ`k , {θk , Θik−1 } for each cluster as if the other clusters do not Form Θ`k , {θk , Θik−1 } for each cluster as if the other clusters do not
exist. exist.
Without clustering Without clustering
Θ1k−1 , P(Θ1k−1 ) Θ2k−1 , P(Θ2k−1 )
yk1 yk2 yk3 yk1 yk2 yk3
measurements measurements
FA FA
FA NT
FA
Cluster-1 Cluster-2 FA NT
FA
Cluster-1 Cluster-2
FA NT NT FA NT NT
T1
FA
NT
Θ1k−1 , P(Θ1k−1 ) Θ1k−1 , P(Θ1k−1 ) T1
FA
NT
Θ2k−1 , P(Θ2k−1 ) Θ2k−1 , P(Θ2k−1 )
yk1 yk2 yk3 yk1 yk2 yk3
FA FA
FA NT measurements measurements FA NT measurements measurements
Hypotheses
Hypotheses
FA FA
NT NT NT FA FA NT NT NT FA FA
Hypotheses
Hypotheses
FA FA NT FA FA NT
Hypotheses
T1 NT T1 T1 NT T1
Hypotheses
FA NT FA NT
FA NT NT FA NT NT
FA NT T1 FA NT T1
FA FA FA FA
T2 NT NT T2 NT T1 NT NT T1 NT
FA T1
T1 NT FA
T1 NT
FA
FA NT
FA
T1 NT NT
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 13 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 13 / 36
Conceptual MHT Simplifications Conceptual MHT Simplifications
Delete hypotheses with probability below a threshold, γp (e.g., This scheme assumes that Θ22
γp = 0.001) any uncertainty is perfectly
Θ11 Θ32
Deletion Condition: P(Θik ) < γp resolved after N time steps
Hypotheses
It is a general commonsense Θ10 Θ21 Θ42
Keep only the most probable hypotheses with a total probability mass
to choose N ≥ 5 (situation
above a threshold, γc (e.g., γc = 0.99) Θ20 Θ31 Θ52
dependent)
i
X The N last ancestors of each Θ30 Θ41 Θ62
Deletion Condition: P(Θ`kk ) > γc hypothesis must be stored
k=1 Θ51 Θ72
` time
where `k is a sequence such that P(Θ`kk ) ≥ P(Θkk+1 ) 0 1 2 3 4
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 14 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 15 / 36
Case N = 2 Case N = 2
Θ12
Hypotheses
It is a general commonsense Θ10 Θ21 Θ42 It is a general commonsense Θ42
to choose N ≥ 5 (situation to choose N ≥ 5 (situation
Θ20 Θ31 Θ52 Θ31 Θ52
dependent) dependent)
The N last ancestors of each Θ30 Θ41 Θ62 The N last ancestors of each Θ41 Θ62
hypothesis must be stored hypothesis must be stored
Θ51 Θ72
time time
0 1 2 3 4 0 1 2 3 4
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 15 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 15 / 36
Conceptual MHT Simplifications Conceptual MHT Simplifications
Case N = 2 Case N = 2
Hypotheses
Hypotheses
It is a general commonsense Θ42 Θ33 It is a general commonsense Θ42 Θ33
to choose N ≥ 5 (situation to choose N ≥ 5 (situation
Θ31 Θ52 Θ43 Θ31 Θ52 Θ43
dependent) dependent)
The N last ancestors of each Θ41 Θ62 Θ53 The N last ancestors of each Θ41 Θ62 Θ53
hypothesis must be stored hypothesis must be stored
Θ63 Θ63
time time
0 1 2 3 4 0 1 2 3 4
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 15 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 15 / 36
Case N = 2
time
0 1 2 3 4
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 15 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 16 / 36
Conceptual MHT Summary Hypothesis-Based MHT
Attractive method since each hypothesis is Proposed by Cox and Hingorani (1996)
• an alternative representation of reality Generate only the best hypotheses, skip hypotheses that will be
• easily interpreted deleted
Drawback: generating all possible hypotheses only to discarding Use the N-best solutions to the assignment problem (introduced last
(most of) them is inefficient lecture with GNN)
Some hypotheses contain the same track; hence fewer unique tracks • Murty’s method, 1968
than hypotheses Find the Nh -best hypothesis, generating as few unnecessary
Track based methods were popular until an efficient way to implement hypothesis as possible
a hypothesis based MHT was given by Cox and Hingorani (1996) Hypothesis reduction techniques still apply
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 17 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 18 / 36
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 19 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 20 / 36
Hypothesis-Based MHT Assignment Problem Hypothesis-Based MHT Assignment Problem
Murty’s Method
Given the assignment matrix Ai ,
Given an assignment matrix Ai , the Auction algorithm (or similar)
Find the best solution using Auction algorithm.
finds the best assignment in polynomial time
2nd best solution:
Generalizations of this problem to find the N-best solutions:
• Express the 2nd best solution as the solution of a number of best
• Formulate as several best assignment problems
solution assignment problems.
• Solve independently using the Auction algorithm
• Find the solution to each of these problems by Auction.
• Murty’s method
• The solution giving the maximum reward (minimum cost) is the second
best solution.
Repeat the procedure for more solutions
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 21 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 22 / 36
j∈JDi
1 − PDj PGj | {z }
Legacy 3. For j = 1 . . . Nh
| {z }
Assignment dependent 1. For i = 1 . . . Nh
1. For the assignment matrix Ai find the jth best solution Θjik
Find {Θ`k }N h
`=1 that maximizes P(Θ`k |y0:k ). 2. Compute the probability P(Θjik )
Two steps: 3. Update HYP-LIST and PROB-LIST: If the new hypothesis enters the
list, discard the least probable entry
• Obtain the solution from the assignment (Murty’s method)
4. If P(Θjik ) is lower than the lowest probability in PROB-LIST discard Θjik
• Multiply the obtained quantity by previous hypothesis dependent terms
and never use Ai again in subsequent recursions
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 23 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 24 / 36
Track-Based MHT Track-Based MHT
Hypotheses usually contain identical tracks — significantly fewer Tracks at time k, {Tki }N t
i=1
tracks than hypotheses Track scores, Sc(Tki )
Idea: Store tracks, T i , not hypotheses, Θi , over time Form a track tree, not a hypothesis tree
Delete tracks with low scores
Θ1k−1 , P(Θ1k−1 ) Θ2k−1 , P(Θ2k−1 )
old tracks new tracks
yk1 Tk1 yk3
1
Tk−1 NM Tk2
yk1 Tk3
Track List
2
Tk−1 NM Tk4
yk1
2
ŷk|k−1 yk2 Tk5
2
ŷk|k−1
1 1
ŷk|k−1 ŷk|k−1 1
Tk−1 2
1
ŷk|k−1
yk1 Tk6 Tk−1
yk2
yk2 Tk7
yk3 Tk8
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 25 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 26 / 36
Track List
old tracks new tracks 2
Tk−1 NM Tk4
Tk1 Tk2 Tk3 Tk4 Tk5 Tk6 Tk7 Tk8 yk2
Sc(Tkj )
X
yk1 Tk1
Sc(Θik ) = Tk5
Tk1 0 0 0 1 1 0 1 1 1
Tk−1 NM Tk2
Tk2 0 1 1 1 1 1 1 Tkj ∈Θik
yk1 Tk3 yk1 Tk6
Tk3 0 0 0 0 1 1
Track List
2
Tk−1 NM Tk4
Tk4 0 0 1 1 1 Hypothesis probability: yk2 Tk7
yk2 Tk5
Tk5 0 1 0 1
exp Sc(Θik )
Tk6 0 1 1
yk1 Tk6
i yk3 Tk8
yk2 Tk7 P(Θk ) =
Tk7 0 1 1+ N j
P h
yk3 Tk8 j=1 exp Sc(Θk )
Tk8 0
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 27 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 28 / 36
Track-Based MHT Implementation Details Track-Based MHT Summary
z −1
User
Presentation
Logic
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 29 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 30 / 36
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 31 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 32 / 36
Examples ADABTS Summary Concluding Remarks
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 33 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 34 / 36
Samuel S. Blackman.
Multiple hypothesis tracking for multiple target tracking. Ingemar J. Cox, Matthew L. Miller, Roy Danchick, and G. E. Newnam.
IEEE Transactions on Aerospace and Electronic Systems, 19(1):5–18, January 2004. A comparison of two algorithms for determining ranked assignments with application to
multitarget tracking and motion correspondence.
Samuel S. Blackman and Robert Popoli. IEEE Transactions on Aerospace and Electronic Systems, 33(1):295–301, January 1997.
Design and analysis of modern tracking systems.
Artech House radar library. Artech House, Inc, 1999. Roy Danchick and G. E. Newnam.
ISBN 1-5853-006-0. Reformulating Reid’s MHT method with generalised Murty K-best ranked linear
assignment algorithm.
Ingemar J. Cox and Sunita L. Hingorani. IEE Proceedings-F Radar and Sonar Navigation, 153(1):13–22, February 2006.
An efficient implementation of Reid’s multiple hypothesis tracking algorithm and its
evaluation for the purpose of visual tracking. Matthew L. Miller, Harold S. Stone, and Ingemar J. Cox.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(2):138–150, February Optimizing Murty’s ranked assignment method.
1996. IEEE Transactions on Aerospace and Electronic Systems, 33(3):851–862, July 1997.
Ingemar J. Cox and Matthew L. Miller. Donald B. Reid.
On finding ranked assignments with application to multitarget tracking and motion An algorithm for tracking multiple tragets.
correspondence. IEEE Transactions on Automatic Control, 24(6):843–854, December 1979.
IEEE Transactions on Aerospace and Electronic Systems, 31(1):486–489, January 1995.
G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 35 / 36 G. Hendeby (hendeby@isy.liu.se) Target Tracking: Lecture 5 (MHT) December 10, 2014 36 / 36