See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327483478

Distributed Flooding-then-Clustering: A Lazy Networking Approach for

Distributed Multiple Target Tracking

Conference Paper · July 2018

DOI: 10.23919/ICIF.2018.8455759

CITATIONS READS

22 148

3 authors:

Tiancheng Li Juan Manuel Corchado Rodríguez

121 PUBLICATIONS 3,732 CITATIONS

Universidad de Salamanca - AIR Institure
1,104 PUBLICATIONS 25,190 CITATIONS
SEE PROFILE
SEE PROFILE

Huimin Chen
University of New Orleans
98 PUBLICATIONS 2,214 CITATIONS

SEE PROFILE

All content following this page was uploaded by Tiancheng Li on 10 September 2018.

The user has requested enhancement of the downloaded file.

 2018 21st International Conference on Information Fusion (FUSION)

Distributed Flooding-then-Clustering: A Lazy Networking Approach for

Distributed Multiple Target Tracking
Tiancheng Li, Juan M Corchado Huimin Chen
BISITE Digital Innovation Hub Dept. Electrical Engineering
University of Salamanca University of New Orleans
Salamanca 37007, Spain LA 70148, USA
{t.c.li, corchado}@usal.es [email protected]

Abstract— We propose a straightforward but efficient net- In this paper, we consider a realistic, challenging scenario
working approach to distributed multi-target tracking, which which faces three particular classes of challenges:
is free of ingenious target model design. We confront two
challenges: One is from the lack of statistical knowledge about • Both the number and statistical models of the latent
the target appearance/disappearance and movement, and about targets regarding either appearance/disappearance or
the sensors, e.g., the rates of clutter and misdetection; The other movement are unknown and may vary over time.
is from the severely limited computing and communication • The sensors may fail to detect a target or report false
capability of the low-powered sensors, which may prevent
data, at unknown (and time-varying perhaps) rates.
them from running a full-fledged tracker/filter. To overcome
these challenges, a flooding-then-clustering (FTC) approach • More challengingly, the low-powered sensors do not
is proposed which comprises two components: a distributed have sufficient computing capability for running a com-
flooding scheme for iteratively sharing the measurements be- plicated multi-target filter, e.g., the sensing rate is much
tween sensors and a clustering-for-filtering approach for target higher than the filter update rate.
detection and position estimation from the local aggregated
measurements. We compare the FTC approach with cutting These challenges frustrate many conventional, ingenious
edge distributed probability hypothesis density (PHD) filters model-driven approaches. Instead of striving to develop a
that are modeled with appropriate statistical knowledge about computationally affordable filter that has to jointly estimate
the target motion and the sensors. A series of simulation studies the targets and the statistics of the sensors (regarding clutter,
using either linear or nonlinear sensors, have been presented
to verify the effectiveness of the FTC approach. detection probability, etc.), we resort to a pure data driven
Index Terms— distributed tracking, flooding, sensor fusion, approach, which is free of complicated model design. Our
average consensus, clustering. approach is analogous to a distributed implementation of the
clustering for filtering approach [17], aided by the distributed
I. I NTRODUCTION flooding scheme for data preparation [18]. In summary, the
protocol is composed of three steps:
It is of both theoretical and practical significance to 1) Disseminate raw measurements iteratively among the
employ a set of geographically dispersed, interconnected neighbor sensors via distributed flooding [18];
sensors, namely a sensor network, to cooperatively handle 2) Convert the locally aggregated measurements to the
a task of high complexity such as tracking a time varying same position space, resulting in a set of point-target-
number of targets [1]–[4]. Without relying on any fusion detections at each sensor;
center, the distributed network formulation enjoys several 3) Cluster the point-detections at each sensor and use the
advantages, including high scalability to adding/removing clusters indicating the number and positions of targets.
nodes and high immunity to local node failure.
In the general distributed target tracking setup, each sensor The proposed flooding-then-clustering (FTC) approach
has sensing, computing, and communication capabilities of will be compared with several distributed probability hy-
its own and runs its own estimator with (iteratively) interact- pothesis density (PHD) filtering [19]–[21] approaches. These
ing with its interconnected neighbors for information sharing. PHD filters are favorably provided with all the required
The shared information can be the measurements (the raw statistical models about the targets and the sensors for the
data [5], [6] or the likelihood function [7]), the prior/posterior best possible performance, erasing the aforementioned com-
distribution [8]–[13], or the estimate/track [14]–[16]. As a puting challenges. That is to say, careful and sophisticated
result, the sensor network progressively reaches consensus model preparation is required before running these filters.
[3], [4] and all sensors benefit from the network cooperation. In contrast, the FTC approach does not need these and is
therefore referred to as a “lazy networking” approach.
J.M. Corchado is also with Department of Electronics, Information and The paper is organized as follows. The flooding and
Communication, Osaka Institute of Technology, Osaka, Japan. clustering approaches are presented in Sections II and III,
This work is in part supported and funded by EU H2020 Marie respectively. The comparison approaches are briefly given in
Skłodowska-Curie Individual Fellowship (Grant no 709267) and MOVI-
URBAN Project (Ref. SA070U 16) co-financed with Junta Castilla y León, section IV. Simulations are presented in Section V before
Consejerı́a de Educacin and FEDER funds. we conclude in Section VI.

978-0-9964527-7-9 ©2018 ISIF 2415

 2018 21st International Conference on Information Fusion (FUSION)

II. P ROBLEM S TATEMENT AND D ISTRIBUTED measurements to its out-neighbors)

M EASUREMENT F LOODING  
Zi (1) = Zi (0) ∪ Zj (0) = Zj (0). (1)
A. Notations and Assumptions j∈Ni j∈Ni (≤1)

The sensor network is represented by a graph with the For t ≥ 2, each sensor collects the new measurements
set of sensors S = {1, 2, · · · , S} and the set of edges E ⊆ that its in-neighbors have received at the preceding iteration
S × S. The edge is denoted by an ordered pair of sensors (while transiting the new measurements it received to its out-
(i, j) ∈ E, which means that sensor j is directly reachable neighbors)
from sensor i, where i is called the in-neighbor of j while   
Zi (t + 1) = Zi (t) ∪ Zj (t) \ Zj (t − 1)
j is the out-neighbor of i. Nj  {i ∈ S|(i, j) ∈ E, i = j} j∈Ni
denotes the set of the in-neighbors of sensor j ∈ S excluding 
sensor j itself. An undirected graph is a special case of a = Zj (0), (2)
directed graph where for any (i, j) ∈ E, we have (j, i) ∈ E. j∈Ni (≤t+1)

Furthermore, we denote the set of all the sensors that are where A \ B is the set difference between A and B, namely,
t ∈ N = {0, 1, 2, ...} hops aways from sensor j as Nj (t) the set of all elements that are members of A but not of B.
t
and that at most t hops aways as Nj (≤ t)  i=0 Nj (i). Given that the network is SC, all sensors will have exactly
Obviously, we have Nj (1) = Nj , Nj (0)  j. the same measurement set when t ≥ Dm , i.e.,
A digraph is said to be strongly connected (SC) if any 
sensor is reachable from all the other sensors. The largest Zi (t ≥ Dm ) = Zj (0). (3)
j∈S
distance between any two sensors, denoted as Dm , is called
the diameter of the graph, which only exists in the SC To monitor the convergence step by step, the degree
network. of consensus (DoC) that can be achieved by the flooding
For the general multi-target tracking in the presence of scheme at iteration t [18] is given as
false and missing data, we make the following assumptions: S
|Ni (≤ t)| − S
(i). Both the number Nk and the states Xk of targets at C o (t) = i=1 ∈ [0, 1]. (4)
S(S − 1)
a particular time k are random. (ii). Each target generates
no more than one at each scan, which is independent of By flooding, C o (0) = 0, C o (t ≥ Dm ) = 1.
other targets. (iii). Clutter is generated at a random rate (e.g., Assuming that each iteration takes time τ , the sensor
Poisson) which is independent of the targets. (iv). Any target scanning period is Δ, and the clustering scheme takes time δ,
may be randomly missed in the detection of a sensor, with the maximum number of iterations allowed without causing
unknown rate. sensor data missing is given by
Δ − δ 
B. Distributed Set-theoretic Flooding tmax = , (5)
τ
In this paper, we employ the distributed flooding approach where the floor operation · finds the largest integer that is
[18] for measurement dissemination over the sensor network, no larger than the argument of the operation. In this paper,
by which all sensors synchronously broadcast their informa- we assume that τ < Δ − δ so that tmax ≥ 1.
tion to out-neighbors. Given that the network is SC, it is able III. DATA - ONLY I NFERENCE AND C LUSTERING
to reach a consensus (i.e., all sensors have exactly the same
set of information) after a certain number of neighborhood A. Observation-only Inference
flooding iterations, in which each sensor aims to “collect We restrict ourselves to the target detection and tracking
data” from all reachable sensors. problem for estimating the number and planar position of
While this communication protocol poses great challenges the latent targets. For clarity, we concentrate on two types of
to the storage and communication capability of the local sensors, with either linear measurement models or nonlinear
sensors, it is quite appealing in certain applications where the measurement models, both of which are made directly on the
sensors have sufficient storage and communication capability. position of the target. Both are “sufficient sensory sensor” for
In short, each sensor stores a table of the initial data of target-position estimation in the plane [17].
all the sensors known at that time. At each communication The linear position measurement model is given as follows
iteration t ∈ {1, 2, . . .}, the neighbors exchange the data in zx,k 1 0 px,k vk,1
their tables (or more precisely, the entries the others do not = + , (6)
zy,k 0 1 py,k vk,2
have), and update their tables. Specifically, the set of the local
measurements each sensor that has Zi = {z1 , z2 , ·, }, i ∈ S where [px,k , py,k ]T gives the target position and vk,1 and vk,2
will be updated to Zi (t) at iteration t ∈ N+ = {1, 2, · · · }. are mutually independent noise.
We use Zi (0)  Zi to denote the initial measurement set at The unbiased projection of the linear measurement to the
sensor i ∈ S. state space is given as follows
To be more specific, when t = 1, each sensor collects p̂x,k zx,k E[vk,1 ]
= − . (7)
measurements from all its in-neighbors (while transiting its p̂y,k zy,k E[vk,2 ]

2416
 2018 21st International Conference on Information Fusion (FUSION)

The nonlinear measurement model we consider is 2) Constraint 2: Over-sized cluster indicates closely dis-
tributed targets and should be divided to meet con-
rk (px,k − mx )2 + (py,k − my )2 vr
= + , (8) straint 1. For this, a distance-based clustering algo-
θk arctan (py,k − my )/(px,k − mx ) vθ
rithm, e.g., the k-means scheme [25] can be applied.
where [mx , my ]T represents the position of the sensor and 3) Constraint 3: A final cluster should not contain more
vrk , vθk are noises affecting the range and bearing measure- than one detection originating from the same sensor.
ments, respectively. In short, we apply a constrained, hierarchical clustering
The projection of the range and bearing to the position scheme that employs the DBSCAN algorithm on the raw
space is (we note here that the formulation given in our data level and the k-means clustering on the finner level.
earlier work [17] is incorrect) Furthermore, inspired by the continuous clustering scheme
[26], a target motion model can be employed to determine
p̂x,k sgn(p̂x,k )bk mx
= + , the potential cluster centroid. That is, the centroids of the
p̂y,k sgn(p̂y,k )tan(|θk − E[vθk ]|)bk my
(9) clusters formed at time k − 1 accumulated by a Markov
 transition - corresponding to the target motion - will form
(rk −E[vrk ])2
where bk  1+(tan(|θk −E[vθ ]|))2 and the signs sgn(p̂x,k ) the potential cluster center at time k. This will speed up the
k

and sgn(p̂y,k ) rely on the quadrant of the target in the computation, while the inaccuracy or error of the Markov
sensing range of the sensor, which is given as follows transition model will not affect the clustering result much.
sgn(p̂x,k ) = 1, sgn(p̂y,k ) = 1 when θk ∈ [0, π/2], In this way, extra consideration has to been given to target
sgn(p̂x,k ) = −1, sgn(p̂y,k ) = 1 when θk ∈ [π/2, π], disappearance and new birth at time k. However, since no
sgn(p̂x,k ) = 1, sgn(p̂y,k ) = −1 when θk ∈ [−π/2, 0] and statistical information is available about the target motion
sgn(p̂x,k ) = −1, sgn(p̂y,k ) = −1 when θk ∈ [−π, −π/2]. a-priori, we do not exploit a time series model to do so in
However, the estimate given by (9) is biased. Given higher this paper. To reduce the communication cost, we may not
order statistical information about the noise, the Monte Carlo apply Constraints 3 (which needs the sensor stamp of each
sampling [17], [22] can be applied to combat the nonlinear measurement to be communicated together with the mea-
projection/conversion bias and to calculate the co-variance surement) but treat it as a soft constraint that does not need
of the estimates. However, as assumed that the higher order to be strictly fulfilled. These simplifications are motivated by
statistical information is unavailable, we ignore it here. the significant consideration of the computational affability
of the low-powered sensors.
B. Measurement Superposition and Clustering
After tmax rounds of the measurement flooding as pre- C. Estimate Extraction from Clusters
sented in Section II-B, the locally aggregated measurements We take the centroid of each cluster as the state estimate of
will all be projected to the target position state space as the corresponding target. When the uncertainties of the point-
addressed in Section III-A, leading to the set of detections estimates in a cluster are obtained by the Monte Carlo ap-
Xs (tmax ), ∀s ∈ S, which include both real target detections proach as addressed and are different, it should be accounted
and false alarms. We note here, one may do measurement for in calculating the weighted centroid of the cluster [17].
conversion prior to the flooding communication. Either way, That is, for a cluster c consisting of target position estimates
[c]
a clustering scheme is next applied to distinguish them and ˆ [c]
{x̂s,k }s=1,··· ,n[c] ⊆ Xs (tmax ), the weighted centroid x̄ s,k is
k
the centroids of the formed clusters are finally extracted as given as
the estimates of the targets’ position. n[c] [c] −1 [c]
The clustering is based on a pivotal statistics about the [c]
k
Cov x̂s,k x̂s,k
ˆ s,k = s=1[c]
x̄ . (10)
distribution of the measurements reported from the condi-  nk [c] −1
s=1 Cov x̂ s,k
tionally independent sensors that observed the same scenario:
the detections are of high density in the area containing Obviously, if these data points are of the same accuracy we
n[c]
targets and are of low density anywhere else. For this task, ˆ [c]
have x̄ 1
s,k = [c]
[c]
s=1 x̂s,k . If the accuracies of these data
k
nk
a straightforward solution is provided by the density-based points are unknown, we also calculate the centroid as this.
clustering method called DBSCAN [23] that distinguishes
Assuming that the measurements between different sen-
the areas of high data density from the low-density regions
sors are conditionally independent, the uncertainty of the
and does not need to be specified with the number of clusters.
fused estimate can be estimated as
To this end, a key parameter regarding the neighborhood
[c]
radius is needed, which is related to the uncertainty of the nk

measurement noise and can be learned from the data if it is Cov ˆ [c]
x̄ s,k = ws (x̄
[c] [c]
ˆ s,k − x̂s,k )(x̄[c] [c]
ˆ s,k − x̂s,k )T , (11)
unknown [24]. This is the only statistical information needed s=1
in our clustering approach. In addition, three constraints can
n[c]
be used to guide the clustering [17], [24] where ws ≥ 0 is the fusing weight (s.t. k
s=1 ws = 1)
[c]
1) Constraint 1: The size of a cluster at sensor s should for which we set uniform weights ws = 1/nk (namely
be between Ns (≤ tmax )/2 and |Ns (≤ tmax )|, cf. (5). assuming identical quality of all point estimates).

2417
 2018 21st International Conference on Information Fusion (FUSION)

D. Algorithm Summary can compensate for the false/missing data by a few sensors.
1) Flooding: share the measurements as many iterations Such an attribute is not obvious with the logarithmic fusion
as possible between neighbors; see Section II-B. [11], [12], which has been proven vulnerable to missing [29]
2) Converting: transfer the locally collected measurement or incorrect data [30].
into the position space; see Section III-A. The AA fusion based distributed GM- and SMC- PHD
3) Clustering: apply the constrained, hierarchical cluster- filters are chosen as comparison approaches [31], [32]. In
ing on the point-estimates; see Section III-B. both approaches, the information communicated between
4) Extraction: calculate the centroids of the clusters as the neighbor sensors is a GM representation of the essential
estimates of the targets’ position; see Section III-C part of the posterior PHD, referred to as “partial PHD”,
whether the local components are GM or particles. They
As we have noted, the order of the first and the second
can cooperate with each other seamlessly in a single sensor
steps may be exchanged.
network, as have been shown in [32]. The consensus achieved
IV. D ISTRIBUTED PHD F ILTERING is called partial consensus [31], [32].
As the goal of reaching a AA consensus, the local PHD
A. PHD Filtering
at sensor s is replaced by the AA of the initial GM-PHDs
The collections of target states at time k can be represented 
GM 1 GM
as a set Xk = {xk,1 , · · · , xk,Nk }, where Nk is the number D̄s,k (x)  Di,k (x), (14)
of targets, which is random. The PHD D(xk ) is defined over |Ns (≤ tmax )|
i∈Ns (≤tmax )
the single target state space and whose integral in any region GM
where Ds,k (x) denotes the GM proxy of the posterior PHD
R ⊆ Rd gives the expected number of targets in that region
at sensor s ∈ S, which is the significant components of the
[19], denoted NkR , i.e.,
 local GM in the GM-PHD filter or is generated from the
D(xk )dxk = E[|Xk ∩ R|] = E[NkR ]. (12) particle distribution in the case of SMC-PHD filter.
R To obtain a crude approximation of D̄kGM (x) at each sen-
Two practical implementations of the PHD filter, namely, sor, the aforementioned flooding scheme given in Section II
the Gaussian Mixture (GM) PHD filter and the Sequential can be applied here (i.e., the flooding items are the GM
GM
Monte Carlo (SMC) PHD filter, represent the posterior PHD parameters corresponding to Ds,k (x)). To reduce communi-
Dk (x|Zs,1:k ) conditioned on the measurement series Zs,1:k cation and memory cost, mixture reduction such as Gaussian
received at sensor s ∈ S up to time k by a GM (see e.g., mixture merging [31] may be applied at each iteration. This
[27]) and by a set of particles (see e.g., [28]), respectively. also helps reduce the uncertainty of the mixture distribution.
We omit the detail here. It is worth noting that, the AA fusion of two estimation
pairs {x̂1 , C1 } and {x̂2 , C2 }, regarding the same real state
B. Average Consensus x, will lead to an estimation pair {x̂f , Cf } with greater error
Given the measurement independence between sensors covariance, namely Cf − Ci is positive semi-definite, where
Zs,1:k , s ∈ S, it is interesting to calculate the averaged i = 1, 2. However, this does not mean that the mean square
estimate of the number of targets over any region R as error (MSE) E[(x − x̂f )(x − x̂f )T ] of the estimate x̂f is larger
than that of either fusing estimates. We have particularly

S
clarified this issue in a separate work [33] in which it has
N̄kR  ωs E[NkR |Zs,1:k ]
also been shown that the GM merging is helpful to reduce
s=1
 the uncertainty without scarifying the MSE.

S
= ωs Ds,k (xk |Zs,1:k )dxk C. Iterated-Corrector Multi-sensor PHD Filter
R

s=1
Based on the same measurement-flooding scheme given in
= D̄k (xk |Z1:S,1:k )dxk . (13) Section II-B, we will also implement the “iterated-corrector”
R (IC) multi-sensor PHD filter [20] for comparison, which
Here, the arithmeticaverage (AA) of the PHDsis defined as basically iterates the PHD updating step by using the mea-
S S
D̄k (xk |Z1:S,1:k )  s=1 ωs Ds,k (xk |Zs,1:k ), s=1 ωs = 1 surement set from each sensor in a random order.
are the averaging weights. The above derivation implies that
V. S IMULATIONS
the AA-PHD fusion (first attempt for which was presented
in [21]) implies imposing the average consensus [3] on the A. Simulation Setup
estimated number of targets between different sensors, for Our simulation setup is quite ordinary in the literature, us-
which the average is supposed to be more robust than the ing a planar region [−1000, 1000]m × [−1000, 1000]m. Two
local estimate. To understand this, we recall an important scenarios were considered with different network topologies
assumption with the PHD filter: both misdetection and clutter and target trajectories. One consisted of 20 sensors (Dm = 4)
are random and are independent between sensors. It is with purely either linear or nonlinear measurement model as
unlikely for the same target to be missed in detection by the shown in Fig. 1 while the other consisted of 18 sensors (a
majority of all sensors. Neither it is likely that a false alarm hybrid of both linear and nonlinear sensors, Dm = 5) as
occur in the same small area to the majority. So, the AA shown in Fig. 4.

2418
 2018 21st International Conference on Information Fusion (FUSION)

1000
Target7: k [47, 49]
We assumed a constant target detection probability pD (xk ) =
Target1: k [5, 40]
800 0.95 for each sensor and uniformly distributed clutter over
600 Target6: k [44, 60]
Target3: k [27, 60] the sensing range of each sensor with an average rate of 10
400
points per scan, implying a clutter intensity κk = 10/20002 .
These parameters were needed by the GM-PHD filter [31],
y coordinate (m)

200 Target5: k [34, 60]

0
but not by the proposed FTC approach.
-200 For mixture reduction regarding to the GM: GCs with a
weight lower than 10−4 were pruned, any two GCs closer
Target4: k [31, 59]
-400

-600 Target2: k [5, 34] than Mahalanobis-distance τ = 4 were merged and the
-800
maximum number of significant GCs to be transmitted and
Target8: k [54, 60] owned by a sensor was 100. Following the idea of “partial
-1000
-1000 -500 0 500 1000 consensus” [31], only GCs having weight larger than Tc =
x coordinate (m)
0.1 were transmitted between neighbors.
Fig. 1. Scenario 1: each target starts at ’’ and ends at ’’ and the links When t = 4, the Network OSPA, average cardinality
between neighbor sensors are given in black dash lines. estimate and average computing time of each iteration step
were given in the upper row of Fig. 2, separately. For
different numbers of flooding iterations for neighborhood
The states of new-born targets werea Poisson random set
3 communication from t = 0 (namely no communication
with the intensity function γk (x) = i=1 λi G(x; mi , Qr ),
between sensors) to t = 4, the time-averaged network OSPA
with Poisson rates λ1 = λ2 = λ3 = 0.05 indicating
and network communication of different approaches were
the expected numbers of new targets to appear at each
given in the bottom row of Fig. 2, separately. We have the
local area and the Gaussian means m1 = [0, 0, 950, −30]T ,
following key findings with regard to the filtering accuracy,
m2 = [−100, 10, −800, 30]T , m3 = [−800, 20, −500, 0]T ,
communication and computation cost, respectively:
and covariance Qr = diag([100, 25, 100, 25]T ) indicating the
initial target state if a new target appears. Each target had a 1) On filtering accuracy:
time-constant survival probability 0.98 and the survival target
• All cooperative filters improved the filtering accuracy
followed a nearly constant velocity motion as given in [31],
by reducing the filter OSPA as compared to the non-
[32] with the state transition noise uk ∼ G(02 , 25I2 ).
cooperative filter. The OSPA reduction basically in-
The above statistical information about the targets were
creased with that of the number of the flooding iter-
favorably provided to the PHD filters for their best possible
ations, till to a convergent/consensual level.
performance, but not necessarily to the proposed FTC ap-
• The arithmetic average fusion approach outperforms
proach.1 Each scenario was simulated with 100 runs using
significantly the IC approach that is sensitive to the sen-
randomly generated measurement series but the same target
sor order and the performance is significantly dominated
trajectories. Each run consisted of 60 filtering iterations. The
by the last sensor. We noticed that, this problem can be
optimal sub-pattern assignment (OSPA) metric [34] was used
mitigated in the approximate product multisensor PHD
to evaluate the position estimation accuracy with the cut-
filter [5], which, however, is computationally costly and
off parameter c = 1000 and the order parameter p = 2.
leads to a new problem called scale unbalance [35] and
Furthermore, we defined [31], [32] Network OSPA as the
the problem worsens as the number of sensors increases.
average of OSPAs obtained by all sensors in the network at
• When t = 0, the performance of the FTC approach was
each sampling step and Time-average Network OSPA as the
indeed poor as a single sensor could not distinguish
average of the Network OSPAs over all filtering steps.
the real target measurement from clutter at all and
To evaluate the communication cost, a GC that consisted
could declare mis-detection easily if any. However, once
of a weight parameter (1 tuple), a 4-dimensional vector
t ≥ 1 namely neighboring sensors communicated, it
mean (4 tuples), and a 4×4 symmetric matrix covariance
outperformed the local non-cooperative PHD filter.
(10 tuples) took 15 tuples and the scale-valued cardinality
• When t ≥ 2, the FTC approach performed super,
parameter took 1 tuple. Each weighted particle took 5 tuples
outperforming all PHD filters. We conjecture that there
(4 for the state vector and 1 for the weight).
are two key reasons contributing to its superiority:
B. Linear Sensor Network 1) the raw measurements were conditionally indepen-
The network topology and trajectories of the targets were dent between sensors;
given in Fig. 1 where each sensor had a linear measurement 2) the real target measurements were unbiased (as we
model as in (6) with vk,1 ∼ G(0, 100) and vk,2 ∼ G(0, 100). assume the noises with zero mean) while the filter
estimates did not necessarily guarantee this; in fact
1 We note that, if the target state transition is not correctly modeled, e.g., the estimates yielded at different sensors might be
a wrong dynamic motion model is chosen or there is an unknown, constant correlated (due to communication) and biased even
input/disturbance, the filters will perform worse (even fail) whereas this
does not matter the performance of the proposed FTC approach as it does the measurements were conditionally independent
not rely on the target dynamics. and unbiased.

2419
 2018 21st International Conference on Information Fusion (FUSION)

Average Comput. Time (s)

300 6 100

Network Cardinality
Network OSPA (m) 200 4 10-1

100 2 10-2

0 0 10-3
0 20 40 60 0 20 40 60 0 20 40 60
Time-average Network OSPA (m)

Filtering Step Filtering Step Filtering Step

200 105

Communications
150 Non-cooperative GM-PHD Filter
Iterated-Corrector GM-PHD Filter
100 104 Average Consensus GM-PHD Filter
Flooding-then-clustering
50

0 103
0 1 2 3 4 1 2 3 4
No. P2P Comm. Iterations No. P2P Comm. Iterations

Fig. 2. Linear sensor network: Average performance of different multi-target detection approaches over the filtering time (when t = Dm = 4, in the
upper row) and over different numbers of flooding iterations (in the bottom row).

2) On communication: The FTC and the IC were the recovered from the disseminated GM by a direct re-sampling
same costly in communication which were significantly less procedure; see the detail given in the Section IV.A.3 of [32].
than the GM-parameters flooding approach applied in the The estimates were extracted from the significant GM in
arithmetic consensus protocol. all PHD filtering protocols except in IC PHD filter, where
3) On computation: the particle set was updated iteratively using the measure-
• Overall, the computing speed of the FTC approach was ments from different sensors which is incompatible with the
very close to that of the local non-cooperative GM-PHD particle-to-GM extraction. Instead, we applied the k-means
filter – actually it is much faster when t < 4. This is clustering [28] to the particle set for estimation extraction in
a key advantage of the FTC approach that is free of the IC protocol, which, however has been demonstrably of
running a filter at local sensors. lower accuracy than the particle-GM extraction [37].
• The multisensor IC PHD filter needs to iterate the The simulation results were given in Figs. 3, which were
updating step and is the most computing costly. highly consistent with those shown in the last simulation (in
Fig. 2), except for two differences:
C. Nonlinear Sensor Network 1) The computing time of the FTC approach (based
In this case, the network topology and target trajectories on nonlinear conversion using (9) which was more
were the same as used in the last simulation but the sensors complicated than that of linear conversion (7) used in
were nonlinear. The FOV (field of view) of each nonlinear the last simulation) was higher than that of the local
sensor was a disc of radius 3000m around the location of non-cooperative SMC-PHD filter that assigned only
each sensor. The measurement model was given as in (8) 200 particles for each potential target.
with vrk ∼ G(; 0, σr2 ), vθk ∼ G(; 0, σθ2 ), where σr = 10m, 2) The average OSPA of the FTC approach and the AA
σθ = π/90 rad/s. The sensor located at [ma,x , ma,y ] had fusion based SMC-PHD filter converged to almost the
a target detection probability pD (xk ) = 0.95G([|px,k − same, highly accurate level, withe the increase of the
ma,x |, |py,k − ma,y |]T ; 0, 60002 I2 )/G(0; 0, 60002 I2 ). Clutter number of flooding iterations t.
was uniformly distributed over the FOV with an average rate 3) When t = 0, the IC protocol performs worse than the
of 10 points per scan, namely κk = 10/(3000 × 2π). other distributed PHD filters because of the poor k-
The local PHD filter was a SMC PHD filter [28], [32], for means clustering on particles for estimate extraction.
which 200 particles were assigned for each potential target
during the resampling scheme [36] at each filtering iteration. D. Hybrid Sensor Network
It is worth noting that strategies such as roughening [36] In this case, we studied a hybrid sensor network consisting
which adds a small zero-mean random variable to the state of both linear and nonlinear sensors with a new network
of each resampled particle, are provably useful for increasing topology (Dm = 5) and new target ground truths, as shown
the diversity of particles after resampling and is adopted in Fig. 4. The models and parameters of both types of sensors
in our implementations. We used the same particle-to-GM were the same as given in the last two simulations. The FTC
conversion scheme as given in [32] but the particles were approach worked with projecting both types of measurements

2420
 2018 21st International Conference on Information Fusion (FUSION)

Average Comput. Time (s)

300 6

Network Cardinality
Time-average Network OSPA (m) Network OSPA (m) 200 4
100

100 2

0 0
0 20 40 60 0 20 40 60 0 20 40 60
Filtering Step Filtering Step Filtering Step
300 105

Communications
Non-cooperative SMC-PHD Filter
200 Iterated-Corrector SMC-PHD Filter
Average Consensus SMC-PHD Filter
104
Flooding-then-clustering
100

0 103
0 1 2 3 4 1 2 3 4
No. Flooding Iterations No. Flooding Iterations

Fig. 3. Nonlinear sensor network: Average performance of different distributed multitarget detection approaches against the filtering time (when the
number of flooding iterations t = Dm = 4, in the upper row) and against different ts (in the bottom row).

1000 The source codes and data are available in the following
800 URL: sites.google.com/site/tianchengli85/matlab-codes/c4f.
600

400 VI. C ONCLUSION

Target5: k [28, 60]
y coordinate (m)

200 Target3: k [22, 60]

We presented a FTC approach for multitarget tracking in
0
the presence of significant false/missing data by using a set of
-200 Target7: k [43, 60] Target4: k [24, 60] geographically dispersed, interconnected, low-powered sen-
-400
Target1: k [4, 26]
sors. Neither any model assumption on the target motion nor
-600 Target2: k [16, 43]
Target8: k [48, 60]
any statistical information about the targets and the sensors
-800 Linear sensors

Target6: k [32, 54]

Nonlinear sensors are needed and therefore, the approach is quite flexible and
-1000
-1000 -500 0 500 1000 robust against different scenarios. It was compared with the
x coordinate (m)
cutting-edge distributed PHD filters with the favorable a-
Fig. 4. Scenario 2 consisting of totally 8 targets and 18 sensors (a half priori statistical models about the targets and the sensors,
are linear sensors and the other half are nonlinear sensors). based on common simulation setups. The results showed that
the simple FTC approach that yields position estimation only
could outperform the well modeled filters, in addition to its
into the same state space. For comparison, each linear sensor advantages in computing and communicating efficiency and
ran a GM-PHD filter [31] and each nonlinear sensor ran a in resistance to background change.
SMC-PHD filter [32], which cooperated seamlessly based on
the GM dissemination and fusion. R EFERENCES
The simulation results were given in Fig. 5. Notably, [1] Y. Bar-Shalom and X. Li, Multitarget-multisensor Tracking: Principles
the FTC approach demonstrated again the superiority lower and Techniques. Yaakov Bar-Shalom, 1995.
computing (than any distributed filters) and higher accuracy [2] C.-Y. Chong and S. P. Kumar, “Sensor networks: evolution, opportu-
nities, and challenges,” Proceedings of the IEEE, vol. 91, no. 8, pp.
(when t ≥ 2), although it did not require/utilize any a- 1247–1256, Aug 2003.
priori model information about the targets and even the [3] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and
sensors. This sounds perversely counter-intuitive as an ex- cooperation in networked multi-agent systems,” Proc. IEEE, vol. 95,
tremely simple algorithm outperforms the carefully-designed, no. 1, pp. 215–233, Jan 2007.
[4] A. H. Sayed, “Adaptive networks,” Proceedings of the IEEE, vol. 102,
ingenious algorithms. However, we note that the practical no. 4, pp. 460–497, April 2014.
filters are inevitably based on (either modeling or calcula- [5] R. Mahler, “Approximate multisensor CPHD and PHD filters,” in 2010
tion) approximation, whose accuracy does not guarantee to 13th Int. Conf. Inf. Fusion, July 2010, pp. 1–8.
[6] B. Khaleghi, A. Khamis, F. O. Karray, and S. N. Razavi, “Multisensor
“consistently” improve with the increase of the number of data fusion: A review of the state-of-the-art,” Information Fusion,
(conditionally independent, unbiased) sensors used [38] as vol. 14, no. 1, pp. 28 – 44, 2013.
the FTC approach does [17]. These being said, we note that, [7] O. Hlinka, O. Sluciak, F. Hlawatsch, P. M. Djuric, and M. Rupp,
“Likelihood consensus and its application to distributed particle filter-
the FTC approach provided no velocity information directly. ing,” IEEE Trans. Signal Process., vol. 60, no. 8, pp. 4334–4349, Aug
Labeling may be considered to obtain integrated track. 2012.

2421
 2018 21st International Conference on Information Fusion (FUSION)

Aver. Comput. Time (s)

300 8 0.8

Network Cardinality
Time-average Network OSPA (m) Network OSPA (m)
6 0.6
200
4 0.4
100
2 0.2

0 0 0
0 20 40 60 0 20 40 60 0 20 40 60
Filtering Step Filtering Step Filtering Step
300 105

Communications
200 Non-cooperative Hybrid-PHD Filter
Average Consensus Hybrid-PHD Filter
104
Flooding-then-clustering
100

0 103
0 1 2 3 4 5 1 2 3 4 5
No. Flooding Iterations No. Flooding Iterations

Fig. 5. Hybrid sensor network: average performance of different multi-target detection approaches over the filtering time (when t = Dm = 5, in the
upper row) and against different ts (in the bottom row).

[8] D. Clark, S. Julier, R. Mahler, and B. Ristic, “Robust multi-object Knowledge Discovery and Data Mining, ser. KDD’96. AAAI Press,
sensor fusion with unknown correlations,” in Sensor Signal Processing 1996, pp. 226–231.
for Defence (SSPD 2010), Sept 2010, pp. 1–5. [24] T. Li, F. D. la Prieta Pintado, J. M. Corchado, and J. Bajo, “Multi-
[9] K. Chang, C. Y. Chong, and S. Mori, “Analytical and computational source homogeneous data clustering for multi-target detection from
evaluation of scalable distributed fusion algorithms,” IEEE Trans. cluttered background with misdetection,” Applied Soft Computing,
Aerosp. Electron. Syst., vol. 46, no. 4, pp. 2022–2034, Oct 2010. vol. 60, pp. 436 – 446, 2017.
[10] R. P. S. Mahler, “Toward a theoretical foundation for distributed [25] A. K. Jain, “Data clustering: 50 years beyond K-means,” Pattern
fusion,” in Distributed Data Fusion for Network-Centric Operations, Recognition Letters, vol. 31, no. 8, pp. 651 – 666, 2010.
D. Hall, C.-Y. Chong, J. Llinas, and M. L. II, Eds. CRC Press, 2012, [26] C. S. Jensen, D. Lin, and B. C. Ooi, “Continuous clustering of moving
pp. 199–224. objects,” IEEE Transactions on Knowledge and Data Engineering,
[11] M. Üney, D. E. Clark, and S. J. Julier, “Distributed fusion of PHD vol. 19, no. 9, pp. 1161–1174, Sept 2007.
filters via exponential mixture densities,” IEEE J. Sel. Topics Signal [27] B. N. Vo and W. K. Ma, “The Gaussian mixture probability hypothesis
Process., vol. 7, no. 3, pp. 521–531, June 2013. density filter,” IEEE Trans. Signal Process., vol. 54, no. 11, pp. 4091–
[12] G. Battistelli, L. Chisci, C. Fantacci, A. Farina, and A. Graziano, 4104, Nov 2006.
“Consensus CPHD filter for distributed multitarget tracking,” IEEE [28] B. N. Vo, S. Singh, and A. Doucet, “Sequential Monte Carlo methods
J. Sel. Topics Signal Process., vol. 7, no. 3, pp. 508–520, June 2013. for multitarget filtering with random finite sets,” IEEE Trans. Aerosp.
[13] F. Meyer, P. Braca, P. Willett, and F. Hlawatsch, “A scalable algorithm Electron. Syst., vol. 41, no. 4, pp. 1224–1245, Oct 2005.
for tracking an unknown number of targets using multiple sensors,” [29] J. Y. Yu, M. Coates, and M. Rabbat, “Distributed multi-sensor CPHD
IEEE Trans. Signal Process., vol. 65, no. 13, pp. 3478–3493, July filter using pairwise gossiping,” in 41th IEEE International Conference
2017. on Acoustics, Speech and Signal Processing (ICASSP), March 2016,
[14] S. Mori, K. C. Chang, and C. Y. Chong, “Comparison of track pp. 3176–3180.
fusion rules and track association metrics,” in 2012 15th International [30] I. Hwang, K. Roy, H. Balakrishnan, and C. Tomlin, “A distributed
Conference on Information Fusion, July 2012, pp. 1996–2003. multiple-target identity management algorithm in sensor networks,” in
[15] B. Noack, M. Reinhardt, and U. D. Hanebeck, “On nonlinear track-to- 43rd IEEE Conference on Decision and Control, 2004, pp. 728–734.
track fusion with Gaussian mixtures,” in 17th International Conference [31] T. Li, J. Corchado, and S. Sun, “Partial consensus and conser-
on Information Fusion (FUSION), July 2014, pp. 1–8. vative fusion of Gaussian mixtures for distributed PHD fusion,”
[16] H. Zhu, M. Wang, K. V. Yuen, and H. Leung, “Track-to-track arXiv:1711.10783 [cs.SY], Nov. 2017, revision is under review in
association by coherent point drift,” IEEE Signal Processing Letters, IEEE Trans. Aero. Ele. Syst.
vol. 24, no. 5, pp. 643–647, May 2017. [32] T. Li, “Distributed SMC-PHD fusion for partial, arithmetic average
[17] T. Li, J. M. Corchado, S. Sun, and J. Bajo, “Clustering for filtering: consensus,” arXiv:1712.06128 [cs.SY], Dec. 2017.
Multi-object detection and estimation using multiple/massive sensors,” [33] T. Li, H. Fan, J. M. Corchado, and S. Sun, “First and second order
Information Sciences, vol. 388-389, pp. 172 – 190, 2017. statistics analysis of arithmetic and geometric average fusion,” to
[18] T. Li, J. Corchado, and J. Prieto, “Convergence of distributed flooding Appear.
and its application for distributed Bayesian filtering,” IEEE Trans. [34] D. Schuhmacher, B. T. Vo, and B. N. Vo, “A consistent metric for
Signal Inf. Process. Net., vol. 3, no. 3, pp. 580–591, 2017. performance evaluation of multi-object filters,” IEEE Trans. Signal
[19] R. P. S. Mahler, “Multitarget Bayes filtering via first-order multitarget Process., vol. 56, no. 8, pp. 3447–3457, Aug 2008.
moments,” IEEE Trans. Aerosp. Electron. Syst., vol. 39, no. 4, pp. [35] C. Ouyang and H. Ji, “Scale unbalance problem in product multisensor
1152–1178, Oct 2003. phd filter,” Electronics Letters, vol. 47, no. 22, pp. 1247–1249, Oct
[20] R. Mahler, “The multisensor PHD filter: II. erroneous solution via 2011.
Poisson magic,” in Proc.SPIE, vol. 7336, 2009, pp. 7336 – 12. [36] T. Li, M. Bolić, and P. M. Djurić, “Resampling methods for particle
[21] R. L. Streit, “Multisensor multitarget intensity filter,” in 11th Int. Conf. filtering: Classification, implementation, and strategies,” IEEE Signal
Inf. Fusion, June 2008, pp. 1–8. Process. Mag., vol. 32, no. 3, pp. 70–86, May 2015.
[22] T. Li, J. M. Corchado, J. Bajo, S. Sun, and J. F. D. Paz, “Effectiveness [37] T. Li, J. M. Corchado, S. Sun, and H. Fan, “Multi-EAP: Extended EAP
of Bayesian filters: An information fusion perspective,” Information for multi-estimate extraction for SMC-PHD filter,” Chinese Journal of
Sciences, vol. 329, pp. 670 – 689, 2016. Aeronautics, vol. 30, no. 1, pp. 368 – 379, 2017.
[23] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based [38] S. Coraluppi, M. Guerriero, P. Willett, and C. Carthel, “Fuse-before-
algorithm for discovering clusters in large spatial databases with track in large sensor networks,” J. Adv. Inform. Fusion, vol. 5, no. 1,
noise,” in Proceedings of the Second International Conference on pp. 18–31, 2010.

2422

View publication stats