Professional Documents
Culture Documents
2 FEB 2017
1
AbstractThis paper presents an information-gain-based sen- tracking performance. There are two main preprocessing steps:
sor selection approach as well as a sensor sensing probability 1) sensor selection and 2) calibration.
model-based calibration process for multihuman tracking in Sensor selection is an important problem in human track-
distributed binary pyroelectric infrared sensor networks. This
research includes three contributions: 1) choose the subset of ing based on distributed sensor networks. As sensors become
sensors that can maximize the mutual information between sen- ubiquitous, coupled with the increasing demand for a shorter
sors and targets; 2) find the sensor sensing probability model to processing time and less resources, it becomes important to
represent the sensing space for sensor calibration; and 3) pro- perform selective fusion so that the decision can be made in
vide a factor graph-based message passing scheme for distributed a timely and efficient manner. The need for sensor selection
tracking. Our approach can find the solution for sensor selection
to optimize the performance of tracking. The sensing probability is further demonstrated by the burden of an increasingly large
model is efficiently optimized through the calibration process in volume of sensory data and by the variability of sensor oper-
order to update the parameters of sensor positions and rotations. ations over time and location. We can reduce the amount of
An application for mobile calibration and tracking is developed. data and improve fusion accuracy by selecting proper sensors
Simulation and experimental results are provided to validate the in a certain location at a certain time.
proposed framework.
Calibration is another important issue in distributed track-
Index TermsBinary sensor networks, calibration, factor ing sensor networks. The tracking algorithms need to read
graph, information gain, matrix factorization, message passing, the geometric parameters of sensors: position and orientation.
multiple human tracking, sensor selection.
The estimation errors of sensor parameters result in degraded
tracking performance. This problem is even worse in multi-
ple target tracking, where the errors of one tracker propagate
into other trackers through the data-to-target association pro-
I. I NTRODUCTION cess [7]. Therefore, a successful distributed tracking system
NDOOR multihuman tracking has many applications requires a calibration component that can estimate the sensor
I in surveillance, security, healthcare, and energy-efficient
buildings [1][4]. Recently, binary sensing technologies have
geometric parameters accurately.
This paper presents a framework for distributed tracking
enabled the development of a distributed binary sensor system with dynamic sensor selection and calibration using binary
for target tracking and recognition [5], [6]. The advantages of pyroelectric infrared (PIR) sensors. The contributions of this
a distributed binary tracking system include: 1) system scal- paper include two aspects.
ability to a large number of sensors; 2) system robustness to 1) Information-Gain-Based Sensor Selection: We proposed
local failures; and 3) a system that is low cost with small an information-gain-based sensor selection method for
sensor communication overhead. selecting sensors that can maximize the mutual infor-
In a complete sensor-based tracking system, typically, pre- mation between sensors and targets. The multiple target
processing is needed before tracking in order to achieve higher information gain is derived and formulated. Given the
localizations of targets, we prove that subregion sensor
selection is better than the whole-space sensor selection
Manuscript received April 7, 2015; revised July 27, 2015; accepted by using less sensors with similar tracking errors.
November 2, 2015. This paper was recommended by Associate 2) Sensing Probability Model for Calibration: The calibra-
Editor M.-Y. Lee. tion method updates the probability model by using the
J. Lu is with the Department of Computer Engineering, University of
Houston at Clear Lake, Houston, TX 77058 USA (e-mail: luj@uhcl.edu). sparse measurements. A mobile application is developed
T. Zhang and F. Hu are with the Department of Electrical and Computer to customize calibration paths. It allows the users to do
Engineering, University of Alabama, Tuscaloosa, AL 35487 USA (e-mail: calibration without covering the entire sampling space.
tzhang21@crimson.ua.edu; fei@eng.ua.edu).
Q. Hao is with the Department of Electrical Engineering, South University The rest of this paper is organized as follows. Section II
of Science and Technology of China, Shenzhen, Guangdong 518055, China reviews the related work. Section III provides the big picture
(e-mail: hao.q@sustc.edu.cn). of our preprocessing method as well as the system model.
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. Section IV presents an information-gain-based sensor selec-
Digital Object Identifier 10.1109/TSMC.2016.2523914 tion method. Section V provides the sensing probability model
2168-2216 c 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
and factor analysis-based calibration. In Section VI, we give sensor parameters. In this paper, we use the model-based
the factor graph model-based distributed belief propagation calibration. The probability of sensor sensing model can
framework for multiple target tracking. Section VII provides be updated by gradient descent algorithm. Then, the senor
the simulation and experimental results. Finally, Section VIII parameters can be computed from the updated sensor sensing
concludes this paper. probability model.
The above preprocessing tasks (sensor selection and cali-
bration) prepare us for PIR sensor-based multihuman tracking.
II. R ELATED W ORK We will briefly introduce our tracking scheme here to illus-
Multihuman tracking and recognition have been devel- trate the effects such as reducing tracking errors after sensor
oped for wireless PIR sensor networks [6], [8], [9]. selection and calibration. We will use factor graph to model
Furthermore, human activity learning using context is also our tracking process since factor graph can well represent the
studied [10], [11]. However, no previous studies in PIR sensor message passing process among distributed PIR sensors.
networks clearly present the sensor selection scheme to give Factor graphs are a straightforward generalization of the
the optimal solution. Tanner graph of Wiberg et al. [27], [28]. Tanner [29] intro-
The existing methods for sensor selection can be gener- duced the bipartite graphs to describe families of codes which
ally divided into two groups: 1) the Bayesian-based methods are generalizations of the low-density parity-check (LDPC)
and 2) the information-based methods. The Bayesian-based codes of Gallager [30] and also described the sum-product
sensor selection considers the whole process as an optimiza- algorithm in this setting. Factor graphs take these graph-
tion problem to find the best solution among all possible theoretic models one step further, by applying them to func-
sensor combination. In [12], a Bayesian I-optimal approach tions. From the factor-graph perspective, a Tanner graph for
is used. Giraud and Jouvencel [13] proposed a geometrical a code represents a particular factorization of the characteris-
approach for the sensor selection. Garnett et al. [14] con- tic function of the code. Factor-graph-based tracking has been
structed covariance functions and performed Gaussian process used in many applications. In inertial navigation systems, the
inference over the functions whose domains are a power set. factor graph has been used for information fusion [31]. A mul-
Ni and Pottie [15] used a Bayesian detection approach to tiple hypothesis tracking with factor graph has been proposed
select a subset of sensors. These methods define the selec- in [32]. The factor graph-based association scheme over joint
tion process as probability problems. The sensor selection probabilistic data association has been developed for closely
can be represented as conditional probability given the sen- moving targets in [33]. In this paper, the factor graph is an
sor observations and background information. Most methods undirected graph which is efficiently used in message passing
of information-based sensor selection in tracking problems algorithms.
compute the distribution of target locations and then select
sensors which give the minimum entropy of the target loca-
III. S YSTEM M ODEL
tion distribution [16][18]. In [19], the KullbackLeibler and
Chernoff sensor selection criteria is provided for event detec- A. Binary Sensor System
tion in wireless sensor network. Zhang and Ji [20] presented The low-cost, low-power PIR sensors are used in our
an interesting approach to calculate information gain between binary sensor system. Simple binary bits are generated, which
two sensors. In our research, the information-based method represent the sensors as triggered (1) or not (0), as
is chosen for sensor selection. We only use the binary codes shown in Fig. 1(a). The advantages of binary sensing include:
generated by sensors to estimate the information gain before 1) tolerance of low signal-to-noise ratios and 2) reduction
and after adding sensors. Our method also considers the tar- of data throughput and communication overhead. We have
get locations as prior knowledge for sensor selection to narrow developed pyroelectric sensor nodes for human tracking and
down the coding space. recognition. Each sensor node contains two PIR sensors.
In most sensor networks, the objective functions of Fresnel lens arrays and masks are used to change the field of
the calibration can be constructed through two approaches: view (FOV) of each sensor. As illustrated in Fig. 1(b), the FOV
1) the model-based method or 2) the data-driven method. In the of one sensor node is segmented into many small partitions.
model-based framework, measures of sensor readings can be Each partition is associated with a sequence of binary code,
predicted based on the sensor models. The sensor parameters e.g., 10, 01, and 11. When human subjects walk through
are then self-tuned to reduce the prediction errors [21][23]. the FOV, the sensor node can generate the binary sequence
In the data-driven framework, measures of sensor readings that can be analyzed to find the location, gesture and activity
are correlated with regression functions of sensor parameters. information of targets.
A two-stage process is usually involved: 1) identification of Now, let the M sensors, S = S1 , S2 , . . . , SM , segment the
regression function parameters and 2) fine-tuning of sensor tracking space into many small regions. Each region is asso-
parameters [24][26]. ciated with an M-bit binary code. Fig. 2(a) shows an example
The advantages of data-driven methods include their inde- of space encoding by three distributed binary PIR sensors.
pendence of sensor physical models and the maturity of With such a space encoding scheme, each code measurement
related algorithms. However, those methods usually assume is associated with one or multiple possible regions. Therefore,
slow varying stimulus sources and the existence of well con- a space decoding scheme has to be developed. In [6], we have
ditioned regression functions between sensor readings and used data-to-target association methods for space decoding.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
J(X, U) (3)
C. Calibration
Let the whole system sampling geometry be G(, ), where
Fig. 2. Illustration of (a) space encoding by three distributed binary PIR
G(, ) = g1 (1 , 1 ), g2 (2 , 2 ), . . . , gm (m , m ). It is deter-
sensors and (b) ideal gridded space encoding case with one target tracking. mined by the positions and orientations of the sensors.
With such a sampling geometry, we can achieve tracking for
multiple targets [6].
The measurement decoding converts codewords into the loca- The goal of calibration is to find these parameters that can
tions. This space encoding/decoding theme is used for target perform the best tracking results, i.e., minimizing the track-
tracking. ing errors. In this paper, we propose a new way to achieve
Assume the tracking space is segmented into N small calibration: instead of minimizing the tracking errors between
regions. Then, the binary codes for these N grids are C = the measured and predicted positions, we minimize the error
c1 , c2 , . . . , cN . ci is an M 1 binary string. An approximate between the measured and predicted sampling geometries.
way to represent these small regions are grids. A simple ideal Let the measured sensing probability matrix be V. It can be
example of gridded space encoding with one target tracking approximately represented by two nonnegative matrices, the
by M = 4 sensors is given in Fig. 2(b). This gridded ideal predicted sensing probability basis matrix G RRT + and
space encoding method is used in sensor calibration. weighting matrix W RFR + , that is
V V = WG (4)
B. Sensor Selection
Presence of multiple targets will cause more ambiguities for where the measured sampling probability matrix V(, ) can
such a binary sensor system. In order to optimize the sensor be calculated from G(, ) and the number of basis in G can
deployment to achieve accurate tracking, sensor selection has be given by R < (FT/(F + T)). The initial predicted posi-
to be made. On one hand, multiple sensors have to be triggered tions and orientations of sensors are known to be prior
to track the whole trajectory; on the other hand, we should knowledge.
not trigger too many sensors to avoid high communication The cost function of calibration for the sensor i, in Euclidean
overhead and sensor energy waste. distance form, can be written as
Assume that there are M binary sensors in the sensor net-
work, and the whole tracking space is segmented into N Ei (i , i ) =
Vi (i , i ) Wi Gi
2 +
Wi
+
Gi
(5)
small grids. Also, assume that the states of K targets are
X = (X 1 , X 2 , . . . , X K ). The state can contain the location and where
Fig. 4. Illustration of the sum-product update rule for factor graph. A. Sensor Node Information Measure
When the sensors are deployed in the distributed sensor
network, the observation space is segmented into many small
D. Message Passing
regions similar as shown in Fig. 2(a). Assume the number of
Because our tracking algorithm is based on message pass- regions is N. Generally, m binary sensors in subset U mean that
ing, here we briefly explain its principle. Our message passing each region can be represented as an m-bit binary code. There
algorithm uses a factor graph. Let f x denote the message are 2m 1 codes (without zeroes), Cm = (cm 1 , c2 , . . . , c2m 1 ),
m m
sent from fusion node f to variable node x. Let xf denote p
where cq denotes the qth binary code which is generated by p
the message sent from variable node x to fusion node f . Also, binary sensors. Each code has the probability of detecting one
let ne(x) denote the set of neighbors of a given node x in the target as
factor graph. As shown in Fig. 4, the sum-product message
1
passing algorithm can be illustrated as two expressions i = m,
p Xj |cm i 1, . . . , 2m 1 (9)
ni
xf (x) = uj x (x) (7) where nm m
i is the number of regions associated with code ci ,
uj ne(x)\ f m
and Xj denotes the jth location with code ci . Then, the
f x (x) = f (v1 , . . . , vM ) vi f (vi ) . (8) information of locating one target given m binary sensors is
v1 vM vi ne( f )\x 1
2 m
HU X|Cm = i H X|C = ci
p cm m m
The factor graph-based message passing algorithm for target
i=1
tracking will be discussed in details in Section VI. m 1
2 nm
i
= p cm
i p Xj |cm
i log2 p Xj |ci .
m
(10)
E. System Diagram
i=1 j=1
The whole system diagram is illustrated in Fig. 5. By select-
Here, we denote HU (X|Cm ) = H(X|U).
ing the sensor set U from sensors S, we have the knowledge
For the case of multiple targets tracking, the probability of K
of sampling geometry G(, ). Such a set of sensors can gen-
targets locating in K regions with one of (2m 1) possible
erate the codes for the tracking space. When the targets are
codes is given by
walking through the encoded space, a binary event sequence is
generated. A calibration process will be used to update sensing K
probability models. In the meantime, the distributed message p X 1,...,K cm
i = p X k cm
i (11)
passing algorithm is performed for multiple target tracking. k=1
The target states will be updated during the message passing where we assume K nm i . As the tracking processes
procedure. for all targets are independent, we can rewrite (11) as
In the next few sections, we will provide the details for each p(X 1,...,K |cm
i ) = p(X |ci ) . The conditional entropy for K
1 m K
m 1
2
2) Provide the sensing probability model for calibration, m i
k m
and use nonnegative matrix factorization algorithm to = p c p X c i j i
update the model. i=1 j=1 k=1
K
Algorithm 1: Self-Calibration
Input: Binary measurements B, calibration paths P,
number of sensors m, and sampling geometry
G(, )
Output: ,
1 i 1;
2 while i < m do
3 Initialization: , ;
4 Wi , Gi PredictedProbFunc(P, gi (i , i ));
Fig. 8. Illustration of calibration with parameters. (a) Orientation. /* initialize */
(b) Positions. 5 Vi BinToProbFunc(Bi , P) ;
/* measurement */
6 Vi Wi Gi ; /* prediction */
Then is equal to 1 + 2 . When introducing the sensing 7 ei
Vi Vi
2 +
Wi
+
Gi
;
probability model and the detection ratio, we can rewrite as 8 while ei not convergent do
V GT
1 2 9 Wi Wi Vi GiT ;
= + (16) i i
1 + 1 1 + 2 WT V
10 Gi Gi W Ti V Ti ;
i i
where 1 = (1 /( 1 )) and 2 = (2 /( 2 )). Note 11 end
that (1 /(1 + 1 )) is equal to 1 and (2 /(1 + 2 )) is 12 Vi Wi Gi ; /* reconstruction */
equal to 2 . 13 i , i CalibrationFunc(Vi );
Once we have 1 or 2 and the known positions of g1 or g2 , 14 i + +;
we can calculate the rotation angle i of the sensor. However, 15 end
the position of the sensor Si can be any location on the blue
circle. To calibrate the positions i of sensor Si , we select two
more grids g3 and g4 from both sides of the middle dotted line,
as shown in Fig. 8(b). Then, we have the triangle relationship VI. FACTOR G RAPH -BASED D ISTRIBUTED
M ESSAGE PASSING
2
d1S i
+ d2S
2
i
2d1Si d2Si cos(12 ) = d12
2
(17) To verify the efficiency of the proposed sensor selection
2
d3S i
+ d4S
2
i
2d3Si d4Si cos(34 ) = 2
d34 (18) scheme (refer back to Section IV) as well as the accuracy of
the sensor calibration (Section V), we need to apply them to
where dnSi denotes the distance between grid gn and sensor practical multihuman tracking applications and test the track-
Si , n = (1, . . . , 4). d12 denotes the distance between two grids ing effect to see if the proposed schemes can reduce tracking
g1 and g2 ; 12 is the angle that two grids g1 and g2 and errors. Therefore, in this section, we will provide our factor
sensor Si hold, which is represented as g1 Si g2 and can be graph-based message passing model, which is the critical part
calculated from (16). Then, the sensor Si can be located at the in our tracking scheme.
intersection point of blue and red circles. Because we have
known the positions of the four grids, by using (17) and (18), A. Factor Graph
the positions i of the sensor Si can be computed.
Fig. 9(a) shows the directed graph model-based multiple
target tracking with unknown data association. The hidden
C. Self-Calibration variable Vt is a Ot K measurement-to-tracker association
It is clear that if we know the measured probability matrix matrix, and is a multidimensional (Ot ) one-out-of-K random
Vi of the sensor Si , we can calibrate the parameters of the variable. Pd and Pfa are the probabilities of detection and false
sensor, as shown in Section V-B. In real applications, it is alarms, respectively. The joint probability function of X, Z,
time-consuming to cover all the grids when doing calibration and V is given by
process. Thus, we only use a few measurements in sensing p(X, Z, V) = p(Z|X, V)p(V)p(X)
probability matrix of a sensor. In other words, we can calibrate
T
T
all sensors with a few trajectories. = p(Zt |Xt , Vt ) p(Vt )
We can use a matrix factorization to update the probability t=1 t=1
basis matrix G. The updated weighting matrix and sensing
T1
probability basis matrix can be used as initial matrices in p(Xt+1 |Xt ) p(X1 ) (19)
self-calibration algorithm. They can be obtained by using the t=1
information of predicted initial sampling geometry G(, ) t K j k Vtjk
and calibration paths P. Note that the binary readings can be where p(Zt |Xt , Vt ) = O
j=1 k=1 p(zt |xt ) .
converted into estimated probability measurement matrix V In a distributed PIR sensor system, the whole tracking pro-
based on the calibration paths P. An implementation of the cess can be seen as a message passing process. Therefore, we
self-calibration method is given in Algorithm 1. change the directed graph model into undirected factor graph.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
Fig. 10. Illustration of (a) two layer PIR sensor node (sensor board and iris
Fig. 9. Illustration of (a) directed graph model and (b) undirected factor mote) and a wireless PIR sensor node with two coded masks and (b) PIR
graph model for tracking. sensor system testbed for multiple human tracking.
Fig. 9(b) shows the undirected factor graph of graphical model 2) For the messages flowing into variable nodes Vt and Xt
converted from directed graph model in Fig. 9(a).
Zt ft2 = 1 (31)
Based on (19), the potential functions of joint factors in the
factor graph are ft2 Xt = ft2 Vt ft2 Zt ft2 f 3 Xt (32)
t
Vt
ft1 = (Vt ) = p(Vt ) (20)
ft2 Vt = ft2 Xt ft2 Zt ft2 ft1 Vt . (33)
= (Zt , Xt , Vt ) = p(Zt |Xt , Vt )
ft2 (21)
Vt
p(X1 ), t = 1
ft = (Xt ) =
3
(22) 3) For the message flowing from Xt to Xt+1
1, t>1
(Xt+1 , Xt ) = p(Xt+1 |Xt ). (23) Xt Xt+1 = Xt1 Xt ft2 Xt f 3 Xt . (34)
t
Fig. 11. Illustration of tracking (a) one target case, (b) two targets case, and
(c) three targets case.
Fig. 15. Comparisons of tracking errors and mutual information with respect
to whole-space and subregion sensor selection processes.
Fig. 19. Illustration of the true trajectories. (a) Diagonal path. (b) Circle
path. (c) Random path.
Fig. 18. Illustration of calibration errors of the sensor parameters at 6 oclock
using full coverage calibration.
TABLE I
T RACKING E RRORS A MONG D IFFERENT T RAJECTORIES AND D IFFERENT C ALIBRATION PATHS
Fig. 22. Illustration of tracking resolutions with sampling rates. (a) 11 Hz.
(b) 17 Hz.
[27] N. Wiberg, Codes and decoding on general graphs, Ph.D. dissertation, Fei Hu received the Ph.D. degree in signal pro-
Dept. Electr. Eng., Linkping Univ., Linkping, Sweden, 1996. cessing from Tongji University, Shanghai, China,
[28] N. Wiberg, H.-A. Loeliger, and R. Kotter, Codes and iterative decoding in 1999, and the Ph.D. degree in electrical and
on general graphs, Eur. Trans. Telecommun., vol. 6, no. 5, pp. 513525, computer engineering from Clarkson University,
1995. Potsdam, NY, USA, in 2002.
[29] R. M. Tanner, A recursive approach to low complexity codes, IEEE He is currently a Professor with the Department
Trans. Inf. Theory, vol. 27, no. 5, pp. 533547, Sep. 1981. of Electrical and Computer Engineering, University
[30] R. G. Gallager, Low-density parity-check codes, IEEE IRE Trans. Inf. of Alabama, Tuscaloosa, AL, USA. His research
Theory, vol. 8, no. 1, pp. 2128, Jan. 1962. interests include security: this is about how to over-
[31] V. Indelman, S. Williams, M. Kaess, and F. Dellaert, Factor graph based come different cyber attacks in a complex wireless
incremental smoothing in inertial navigation systems, in Proc. IEEE Int. or wired network. Recently he focuses on cyber-
Conf. Inf. Fusion (FUSION), Singapore, 2012, pp. 21542161. physical system security and medical security issues, signals: this mainly
[32] H. Wang, J. Sun, S. T. Lu, and S. Wei, Factor graph aided multiple refers to intelligent signal processing, that is, using machine learning algo-
hypothesis tracking, Sci. China Inf. Sci., vol. 56, no. 10, pp. 16, 2013. rithms to process sensing signals in a smart way in order to extract patterns,
[33] V. P. Panakkal and R. Velmurugan, Effective data association scheme and sensors: this includes micro-sensor design and wireless sensor networking
for tracking closely moving targets using factor graphs, in Proc. IEEE issues.
Nat. Conf. Commun. (NCC), Bengaluru, India, Jan. 2011, pp. 15.