A. Chen et al.

(2011) “Efficient Boundary Tracking Through Sampling,”

Applied Mathematics Research eXpress, Vol. 2011, No. 2, pp. 182–214
Advance Access publication April 4, 2011
doi:10.1093/amrx/abr002

Efficient Boundary Tracking Through Sampling

Alex Chen1 , Todd Wittman1 , Alexander G. Tartakovsky2 ,

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

and Andrea L. Bertozzi1
1 Department of Mathematics, University of California, Los Angeles, CA
90059, USA and 2 Department of Mathematics, University of Southern
California, Los Angeles, CA 90089, USA

Correspondence to be sent to: Alex Chen, 520 Portola Plaza Math Sciences Building 6363,
Mailcode: 155505, Department of Mathematics, University of California, Los Angeles, CA 90059,
USA. e-mail: [email protected]

The proposed algorithm for image segmentation is inspired by an algorithm for

autonomous environmental boundary tracking. The algorithm relies on a tracker that
traverses a boundary between regions in a sinusoidal-like path. Boundary tracking is
done by efficiently sampling points, resulting in a significant savings in computation
time over many other segmentation methods. For noisy images, the traversed path is
modeled as a change-point detection problem between two states. Change-point detec-
tion algorithms such as Page’s cumulative sum procedure are adapted in conjunction
with other methods to handle a high level of noise. A modification for the multiple-
region case is also presented as a hybrid of a topology-detecting segmentation algorithm
and boundary tracking. Applications to high-resolution images and large data sets such
as hyperspectral are of particular interest. Irregularly shaped boundaries such as frac-
tals are also treated at very fine detail along with accompanying fractal dimension cal-
culations, which follow naturally from the boundary-tracking data.

1 Introduction

Interpreting important features in an image often involves some simplification that dis-
cards detail in favor of generality. In image segmentation, two approaches are partic-
ularly common: region-based methods and edge-based methods. Partitioning an image

Received October 6, 2010; Revised January 24, 2011; Accepted February 24, 2011


c The Author(s) 2011. Published by Oxford University Press. All rights reserved. For permissions,
please e-mail: [email protected].
 Efficient Boundary Tracking Through Sampling 183

into homogeneous regions gives a “cartoon-like” appearance which identifies the most
important objects and their rough boundaries. Many algorithms using this approach
rely on identifying some feature common to each region and classifying individual pix-
els according to how well they match this feature. The classification is often based on
matching pixel intensities with some spatial bias, such as regularity of object bound-
aries. This idea of image segmentation is the method adopted in region-based methods
like the algorithm of Chan and Vese [13], which uses the level set method [47] to mini-

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

mize the piecewise constant Mumford–Shah functional [45]. In many of these methods,
changes in topology are also allowed, so that multiple objects may be detected.
Another approach is to consider the statistical patterns observed in real data.
The seminal work of Geman and Geman [24] describes attributes such as pixel inten-
sity and image edges as a Markov random field. This model places images in a Bayesian
framework, in which features can be described based on similar features found in the
same image or in other data sets that share the same characteristics. Methods such as
region competition rely on this statistical model and are often solved using energy min-
imization [61, 68]. Sometimes the segmentation can be successively refined with further
user interaction [51].
These approaches are generally less sensitive to noise, due to the spatial regular-
ity requirement and the use of information from many pixels in the image. This robust-
ness, however, comes at the expense of computation time, since calculations often must
be done on every pixel, even if some pixels are far from object boundaries or clearly
belong to a certain region.
Recent advances have greatly improved the computational efficiency of segmen-
tation methods based on the Mumford–Shah functional using more efficient methods
for energy minimization [19, 22, 23, 26]. Another possibility is to make calculations only
in a narrow band around the zero level set, reducing the need to make irrelevant calcu-
lations [49].
Alternatively, tracking edges or boundaries of objects gives the location at which
significant changes in the image take place. Local searching algorithms such as image
snakes and gradient vector flow [36, 66] are designed to work faster than region-based
methods. Note that since curves are evolved near the boundaries of objects, it is only
necessary to consider pixels near these boundaries.
In contrast to the region-based methods, noise is often a problem with such
local searching methods since far fewer pixels are used in calculations. Moreover, such
methods often use an edge-detection function based on a gradient. While the gradient
is intended to locate edges, it has the side-effect of making the edge-detector highly
sensitive to noise.
184 A. Chen et al.

Furthermore, certain shapes such as those with large concavities or sharp cor-
ners may not be detected accurately due to requirements in the regularity of the curve.
These points can be somewhat resolved with a balloon term [17] designed to encourage
evolution past local minima of the underlying energy functional or replacing the second-
order partial differential equations (PDE) for snakes evolution with a fourth-order PDE
designed to preserve corners [10, 62]. In [8], a new energy was proposed to replace length
with local area. This allows irregularly shaped objects, especially those with fractal-like

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

structure, to be segmented more accurately. Other edge-based methods include the use
of random walks to test hypotheses for the most likely layout of edges [2] and the Canny
edge detector [11].
Theoretically, the use of fewer pixels in local searching gives these methods a
shorter run time than with region-based methods. These algorithms scale only with the
number of pixels in the boundary, rather than the number of pixels in the image. This
speed advantage is thus even more pronounced for large images.
In this work, we propose a novel segmentation method that seeks to minimize
the number of calculations. By sampling points, the boundary of an object can be
tracked without the need to process a large number of pixels. The method is based
on an algorithm for tracking environmental boundaries [34, 37] which utilizes robots
that “walk” in a sinusoidal path along the boundary between two regions. The robots
change directions as they cross from one region into another. The robot walking method
is related to the class of “bug” algorithms that seek a target while avoiding obsta-
cles [16, 46]. Such algorithms utilize only local information and theoretically should not
visit the same location twice. Only the points that are near the boundary in question are
tracked, resulting in substantial savings in run-time.
Sinusoidal tracking patterns have been observed in a variety of contexts as an
efficient path for travel. Ants walk in a sinusoidal manner along pheromone trails laid
down by other ants, predicting the proper direction of travel and compensating if they
stray too far in a certain direction [18, 30]. For atomic force microscopy, using a sinu-
soidal pattern for scanning can avoid imaging points that are irrelevant to the region of
interest [1, 14].
Some “walking” methods for boundary tracking exist in image processing as
well. Moore’s algorithm [27] tracks an edge of a binary image using an ordered search
for image intensity values of 1 (alternatively, 0) through a neighborhood of each pixel
in the iteration. Another method follows the maximum gradient magnitude of the 3 × 3
neighborhood of each pixel [12]. These methods, however, clearly fail in the presence of
any noise.
 Efficient Boundary Tracking Through Sampling 185

As with local searching methods in image processing, noise can cause problems
since the tracking is done as a local search. It was proposed [34] that the use of a change-
point detection algorithm such as Page’s cumulative sum (CUSUM) procedure [48] would
allow objects to be tracked in noise. Testbed implementations of the boundary-tracking
algorithm suggest that robots can indeed track boundaries efficiently in the presence of
a moderate amount of sensor noise [35].
While the particular algorithms that constitute the basis of the proposed seg-

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

mentation method are not new, their combination is unique and proves to be very effi-
cient for noisy images, in terms of both accuracy (quality of segmentation) and compu-
tational complexity. Furthermore, the setting in image processing poses new problems
as well as benefits. For the discrete image grid, sampling cannot be done on an arbitrar-
ily precise scale. Image resolution is thus a particularly important issue. But an added
benefit is that the entire image can be “seen,” allowing non-local algorithm possibil-
ities. Lastly, the operation on high-dimensional hyperspectral data sets constitutes a
still different application with its own challenges.
One of the greatest advantages for using the algorithm for segmentation is the
computational efficiency. By considering only pixels near the boundary of an object, as
with local searching methods, many pixels are not considered. Furthermore, the track-
ing method travels through each location only once, resulting in run-time savings even
over other local searching methods. As mentioned earlier, noise can be particularly trou-
blesome when fewer points are considered. Further improvements can be made that are
not practical in the environmental tracking case. Many of these improvements are based
on hypothesis testing for two regions, with the use of the CUSUM algorithm as a special
case. Hypothesis testing is an important part of many statistical segmentation meth-
ods [7, 25, 50].
The boundary-tracking algorithm in the context of image processing was briefly
introduced in [15]. In this work, we give a more detailed treatment with expanded the-
ory and applications. In subsequent sections, we will adapt the boundary-tracking
algorithm to the image processing problem and discuss improvements that can be
made in this context. Section 2 briefly reviews the boundary-tracking algorithm and
compares the original environmental tracking and image processing problems. In
Section 3, various methods for better performance in the presence of noise are sug-
gested. Other improvements to the algorithm are discussed in Section 4. Section 5
shows numerical results and images, including examples with hyperspectral and
high-resolution data. Finally, discussion and issues for further study are given in
Section 6.
186 A. Chen et al.

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Fig. 1. (1.1) Global search is initialized at point 1, traveling in a spiral-like pattern. At point 2,
a boundary point is found and local sampling starts tracking the boundary more closely. (1.2)
Basic procedure for the boundary-tracking (local sampling) algorithm. Local sampling works on
a sub-pixel level.

2 A Two-step Locating and Tracking Method

The boundary-tracking algorithm is first initialized with some given decision function,
which determines whether it is in one of two regions. Boundary tracking is done as
a two-step process using the decision function in both steps. The first step involves
locating a boundary point via a global search. The boundary point found by this global
search serves as an initial point to be used in the local sampling step. At this second
step, boundary points are found by using a tracker that travels near the boundary in a
sinusoidal path. Sometimes the tracker can move off the boundary, particularly when
the noise is high or when the decision function is not very accurate. When this happens,
it is necessary to use the global search step again to locate the boundary. This paper
discusses the global search step briefly, and focuses on the local sampling algorithm.

2.1 Global search to locate a boundary point

There are several options for the global search step. The simplest is just a user-defined
point near the boundary that can be directly used by the local sampling step. For a more
automated approach, a random initial point in the image can be given. Then the tracker
travels in a spiral away from the initial point until a boundary point is detected (see
Figure 1). It is sometimes necessary to refine this estimate, since the spiral pattern may
not detect the exact crossing point accurately. Another option is to use a Levy flight to
detect a boundary. Levy flights have been observed in a variety of contexts [39, 64].
 Efficient Boundary Tracking Through Sampling 187

A different approach for global search is to use a hybrid method. Boundary

location can be done by using a coarse segmentation by another method. The result-
ing segmentation gives initial boundary points to be used in the local sampling step.
This method will be explained further in Section 5. For the tracking of one object, how-
ever, global detection is not usually the limiting step in accuracy or in run-time. Thus,
in this work, we focus on the local sampling step.

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

2.2 Local sampling to track the boundary

For the physical problem [34, 37], a robot is used to track an environmental boundary.
The robot is placed near the boundary in question, and it then uses a bang–bang steering
controller to move through the boundary of the two regions.
It is relatively straightforward to adapt the algorithm for images. Let the image
domain be represented by Ω, with B the boundary between two regions Ω1 and Ω2 , so
that Ω = Ω1 ∪ Ω2 ∪ B and Ω1 ∩ Ω2 = ∅. Define an initial starting point x0 = (x01 , x02 ) for the
boundary tracker and an initial value θ0 , representing the angle from the +x1 direction,
so that the initial direction vector is (cos θ0 , sin θ0 ). Also define the step size V and angu-
lar increment ω, which depend on estimates for image resolution and characteristics of
the boundary to be detected. In general, V is chosen smaller for greater detail, and ω is
chosen smaller for straighter boundaries. A decision function between Ω1 and Ω2 must
also be specified and has the following form:
⎧
⎪
⎪ 1 if xΩ1 ,
⎪
⎨
x) = 0
d( if x B, (1)
⎪
⎪
⎪
⎩
−1 if xΩ2 .

The simplest example is thresholding of the image intensity I (

x) at a given spatial loca-
tion x (in the case of a grayscale image):
⎧
⎪
⎪ 1 if I (
x) > T,
⎪
⎨
x) = 0
d( if I (
x) = T, (2)
⎪
⎪
⎪
⎩
−1 if I (
x) < T,

where T is a fixed threshold value. Later in this section, we use statistical information
x). At each step k, the direction
about prior points sampled along the path to modify d(
θk and current location xk are updated recursively. Specifically,

xk = xk−1 + V ∗ (cos θk−1 , sin θk−1 ), (3)

188 A. Chen et al.

and θk is updated according to the new location of the tracker xk. A simple update for θ
is the bang–bang steering controller, defined by

θk = θk−1 + ω d(
xk). (4)

Midpoints of the tracker locations, the iteration before and the iteration after
change, are taken to be the boundary points. Linear interpolation can be used to obtain

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

a boundary curve. Note that unlike many other segmentation methods, it is not assumed
that the curve is smooth. This allows the detection of very irregular objects, especially
those with large concave regions or high curvature boundaries.
An angle-correction modification [34] can be used for (4), if step k is a region
crossing:
θk−1 + d(
xk)(t̄ω − 2θref )
θk = , (5)
2
where t̄ is the number of steps since the last region crossing, and θref is a small fixed
reference angle. The parameter θref is determined empirically with a value of 0.1 giving
good results in many images. Intuitively, θref makes angle correction more conservative
so that the direction of travel is not parallel to the expected direction of the boundary.
This results in some distance still being traversed before the next boundary crossing.
A further consideration is the stopping condition. Several options for the ter-
mination of the algorithm are possible: the tracker completes a certain number of iter-
ations, arrives at the image border, or arrives near the first boundary point detected
with some minimum number of iterations. The latter two were used for numerical
experiments.
There are several differences between the environmental problem and image
processing problem. The primary difference is the conversion from a continuous model
to a discrete model. While a robot samples data wherever it travels, image data are pixe-
lated so that the tracker cannot sample data at the sub-pixel level. Instead, the nearest-
neighbor intensity reading is used. Thus, the level of detail is limited by the image
resolution. However, while sampling is only done at the pixel locations, the tracker still
travels at the sub-pixel level, as indicated in Figure 1. Moreover, the tracker can move
with step sizes less than 1. The redundancy of sampling the same pixel more than once
makes the algorithm more robust to noise and gives a sharper segmentation without
increasing computation time much. Indeed, numerical experiments confirm that the step
size giving optimal results is often less than 1.
Conversion to the image segmentation problem also has several advantages. A
robot is restricted to sampling data only at its current location. In images, however,
 Efficient Boundary Tracking Through Sampling 189

information from the pixels surrounding the tracker can be used to mitigate mistakes
due to noise. Using these extra pixels, of course, negates some computational benefits,
but as long as the number of extra pixels considered is small, the main advantages
are retained. Various methods for using this nonlocal information while preserving the
speed of the algorithm are presented in the next section. Also, robots used in experi-
ments [31, 35] make smooth changes of direction, but this restriction is not required
in the image processing case. Lastly, robots have a fixed position at any point in time

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

and cannot make instantaneous jumps. In images, however, it is possible to place the
tracker in another location in a single step. This ability allows the boundary tracker
to correct mistakes detected at a later time without having to backtrack through many
pixels.

3 Change-Point Decisions

The boundary-tracking method presented works well when a clear and accurate deci-
sion function can be defined. However, in many applications a clear distinction between
the two regions cannot be made, particularly in noisy images. For boundary tracking,
errors in classification by the decision function can lead to serious errors in tracking,
since the local sampling algorithm is only valid when the tracker is near the boundary.
With some method to average nearby pixels, however, the algorithm can be made much
more robust to noise.
Change-point detection theory is well suited to tracking image edges in noise.
In particular, the CUSUM algorithm has been used to improve tracking performance in
the environmental tracking problem [35, 38]. This section reviews relevant information
from change-point detection and its application to the boundary-tracking problem.
Change-point problems deal with rapid detection of abrupt changes in statisti-
cal properties (distributions) of data. One standard application of change-point detec-
tion is in manufacturing [20, 43, 53]. For a certain process, it may be acceptable to have
a certain failure rate. If the process is able to operate below this tolerance level, then
it is allowed to continue operation. If, however, this tolerance level is exceeded, one
would like to stop as soon as possible to make repairs. Making a false stop, how-
ever, is costly as well, so it is important to balance the two considerations. Other
applications include surveillance, computer network security (rapid detection of intru-
sions), failure detection in dynamical systems and communication networks, financial
markets, seismology, navigation, speech segmentation, etc. See, e.g., [9, 59, 60, 65] and
references therein.
190 A. Chen et al.

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Fig. 2. Flowchart of the boundary-tracking algorithm.

More explicitly, given a sequence of independent observations s1 = I (x1 ), . . . , sn =

I (xn) and two probability density functions (pdf) f (pre-change) and g (post-change),
determine whether there exists N such that the pdf of si is f for i < N and g for i ≥ N.
One of the most efficient change-point detection methods is the CUSUM
algorithm proposed by Page in 1954 [48]. Write Z k = log[g(sk)/ f(sk)] for the log-likelihood
ratio and define recursively

Uk = max(Uk−1 + Z k, 0), U0 = 0 (6)

the CUSUM statistic and the corresponding stopping time τ = min{k | Uk ≥ Ū }, where Ū
is a threshold controlling the false alarm rate. Then τ is a time of raising an alarm. In
our applications, assuming that f is the pdf of the data in Ω1 and g is the pdf in Ω2 ,
the value of τ may be interpreted as an estimate of the actual change-point, i.e., the
boundary crossing from Ω1 to Ω2 .
Changes from Ω2 to Ω1 can also be tracked in this manner. Analogously to (6)
define recursively the decision statistic L k = max(L k−1 − Z k, 0), L 0 = 0 and the stopping
time τ = min{k | L k ≥ L̄}, where L̄ is a threshold associated with a given false detection
rate. The description of the basic boundary-tracking algorithm is now complete; the
steps are shown in Figure 2.
The CUSUM algorithm and optimality can be understood intuitively. Consider
two distributions of data D f and Dg with density functions f and g, respectively. In
the change-point detection problem, observations are conditionally independent given
the point of change. Given a sequence of observations xk−l+1 , . . . , xk for a fixed natural
 Efficient Boundary Tracking Through Sampling 191

number l (the change point), test the hypotheses

H0 : xk−l+1 , . . . , xk are all from distribution D f ,

H1 : xk−l+1 , . . . , xk are all from distribution Dg .

g(xk−l+1 )···g(xk )
Then the likelihood ratio test gives the procedure f(xk−l+1 )··· f(xk )
≥ α.
This test, however, requires k − l + 1 observations after a change has actually

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

occurred, since the hypotheses are that the observations all come from one distribution
or the other. Instead, the number of observations l can be allowed to vary, which is
beneficial for our applications where the point of change from D f to Dg is not known
in advance. In this case, a reasonable approach is to declare that a change is in effect
g(xl )···g(xk )
if there exists some l = 1, . . . , k such that f(xl )··· f(xk )
≥ α for some k ≥ 1. In other words,
the change is declared when it is reasonably certain that some number of the most
recent entries provides evidence that a change has occurred. The threshold α describes
the confidence level for the change. This modification allows for the reduction of the
number of false alarms due to noise while still allowing rapid detection in case of an
actual change.
Taking the logarithm, this rule can be associated with the stopping time
  k
 g(xi )

τ = min k  log ≥ Ū for some l = 1, . . . , k
 f(xi )
i=l
 
 k

= min k max Z i ≥ Ū ,
1≤l≤k
i=l

g(xi )
where Ū is a certain threshold that controls the rate of false alarms and Z i = f(xi )
is the
log-likelihood ratio for the ith observation.
k
It is easily seen that the trajectories of the statistic max1≤l≤k i=l Z i coincide
k
with the trajectories of the CUSUM statistic Uk = maxl≥1 i=l Z i on the positive half-
plane. Note that the CUSUM statistic also obeys the recursion (6) (i.e., it is a reflected
from the zero barrier random walk). Therefore, whenever the threshold Ū is positive
(which is usually the case) the stopping time τ is nothing but the CUSUM algorithm that
can be equivalently written as

τ = min{k | Uk ≥ Ū },

where Uk is the CUSUM statistic given in (6).

A few basic observations from (6) can be made. As mentioned earlier, the
algorithm indicates a change from Ω1 to Ω2 only when observations generated by Ω2
192 A. Chen et al.

accumulate. Furthermore, an accumulation of many (negative) values from Ω1 does not

affect performance since the minimum value for U is 0. Lastly, the CUSUM procedure
can be interpreted as a collection of one-sided sequential probability ratio tests (in this
connection, see [40]).
The pdfs f and g for the two regions can be modeled in a variety of ways.
The method is sufficiently general to allow for the use of discriminative or genera-
tive statistical models. One simple generative approach is to assume that f and g are

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

generated by Gaussians or a mixture of Gaussians, a common approach in image pro-
cessing [28, 33, 57]. Another approach is to learn parameters as points are sampled. One
simple way is to update the mean and variance of each region as new points are sam-
pled. In the case of Gaussian noise, this is actually sufficient to describe the entire dis-
tribution. Another slightly more computationally expensive method is to keep a running
update of points classified in each region and allow the resulting histograms themselves
serve as an approximation for the pdfs in each region. Lastly, the entire image could also
be sampled and clustered before the algorithm commences to determine possible his-
tograms. Similar semi-supervised statistical methods have also been seen in [7, 50].
In fact, it is not necessary to have a priori knowledge of the pdfs associated with
different regions. Instead, one can use a “score”-based CUSUM where the log-likelihood
process Z k is replaced with an appropriate score function sk. In our applications, we are
mostly interested in detecting a change in the mean value, in which case the linear score
function is appropriate. To be specific, the following CUSUM statistic may be used:

Uk = max{Uk−1 , 0} + (sk − μ0 − c)σ ,

where σ is the standard deviation of the observations sk (usually estimated approxi-

mately in advance), μ0 is the pre-change mean (known or estimated on-line) and c is
a tuning parameter (usually fixed). If the post-change mean is known, denote it as μ1 .
Then, in the Gaussian case, the optimal value of c is c = (μ1 − μ0 )/2, which gives a formu-
lation equivalent to Equation (6). Even if the pdfs f and g are not known, the algorithm
performs well in practice as long as the signal-to-noise ratio is not too high. Then μ1
is chosen as the minimal mean intensity which can be detected with reasonable perfor-
mance. Note that this statistic does not incorporate pdfs and works for many different
distributions.
In adapting change-point detection theory to region changes, the assumption of
independent observations has been made. The samples are not independent, however,
since they are taken from the tracking path. This assumption is not altogether prob-
lematic in practice, especially when noise is spatially uncorrelated. With a high level of
 Efficient Boundary Tracking Through Sampling 193

noise, the spatially uncorrelated noise gives approximate independence. For a low level
of noise, using change-point detection is less important.
As alluded to earlier in this section, the effectiveness of a change-point detection
algorithm can be quantified. Generally, the important factors are the delay in detection
of a real change (which is random) and false alarms. Detection delay measures the num-
ber of observations it takes to detect a change after it has actually occurred, while a false
alarm occurs when a change is detected but has not actually occurred (Type 1 errors).

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

In general, the detection delay should be small and the false alarm rate should be low.
Clearly, the two goals are antagonistic. By increasing thresholds Ū and L̄ the false alarm
rate can be lowered, but this will unavoidably lead to an increase in the detection delay.
Thus, one has to adjust parameters in the change-point detection algorithm to moderate
both indices.
Typically, operating characteristics of change-point detection algorithms are
expressed via the average run length (ARL) to detection versus the ARL to false alarm.
These performance indices have been first introduced by Page in [48]. The ARL is the
average time to a change detection for a certain scenario. Under this notation, the false
alarm rate is measured by the ARL when no change takes place (i.e., by the mean time
to false detection El (τ |τ < l)), while the average detection delay is measured by the ARL
when a change takes place immediately after observations begin, i.e., E 0 τ . The ARL
is not the only method to measure the effectiveness of change-point detection algo-
rithms [9, 40, 58]. For example, why should we restrict our attention to the detection
delay when the change occurs from the very beginning? The conditional average detec-
tion delay El (τ − l|τ > l) for any fixed point of change l > 0 is a more reasonable measure.
Also, in place of measuring the false alarm rate with the ARL to false alarm one may pre-
fer to work with a probability of false alarm in a fixed time interval. See [58] for a more
detailed discussion.
In 1971, Lorden [40] introduced the minimax (worst-worst case) performance
measure—the essential supremum average detection delay DD = ess sup supl El (τ −
l)+ |x1 , . . . , xl ) and proved that the CUSUM procedure is asymptotically optimal with
respect to this detection delay measure for a low false alarm rate when the ARL to
false alarm is large, among all procedures for which the ARL to false alarm is fixed at
a given level. Later, Moustakides [44] proved that actually CUSUM is exactly optimal for
any false alarm rate with respect to Lorden’s criterion. Other optimality properties of
CUSUM are discussed in [58]. Knowledge of the pdfs is needed in order to prove optimal-
ity results of CUSUM for the change-point detection problem. The adaption of change-
point detection to boundary tracking changes the problem so that the optimality results
194 A. Chen et al.

cannot be applied directly. Nevertheless, the cited results give a guideline as to param-
eter choice and an estimate of effectiveness.
For CUSUM, the following relation holds (asymptotically as ARL to false alarm
is large):
log ARLFA
DD ∼ El (τ − l|τ > l) ∼ , l ≥ 0,
K(g, f)
where ARLFA denotes the ARL to false alarm and

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

g(x)
K(g, f) = E g log
f(x)
is the Kullback–Leibler information number [9]. It can also be shown that asymptotic
detection delay is
Ū
DD ∼ .
K(g, f)
This form of asymptotic detection delay is used to improve boundary tracking in the
following section.

4 Further Improvements to the Boundary-Tracking Algorithm

This section introduces further modifications to the algorithm, possible only in the
image processing case, improving performance in the presence of noise. To mitigate the
effects of noise further, a mean filter can be used; that is, the nearest-neighbor intensity
at the tracker can be replaced by an average of intensity values of nearby pixels. The
idea is that nearby pixels are most likely to be in the same region, so that taking an
average filters noise without using values from the other region. The same principle is
at work with region-based segmentation methods, in which it is assumed that an image
can be partitioned into a few contiguous regions with the same characteristics.
Mean filters can help the performance of the boundary tracker in noise. But
since the tracker follows the boundary closely, it is important to use the uniform region
assumption conservatively. Using a large number of pixels around the tracker would
average noise but likely would also use values from the other region. Using these values
can result in a much more serious error. In practice, a 3 × 3 window centered on the
tracker works well. Instead of averaging the entries in the window, one can apply rea-
soning similar to the hypothesis testing mentioned for the CUSUM algorithm. As with
observations along the boundary-tracking path, the entries in the 3 × 3 window can also
be treated as independent measurements. Thus, in the hypothesis testing of the CUSUM
procedure, the nearest-neighbor observation is replaced by the nine observations in the
3 × 3 window.
 Efficient Boundary Tracking Through Sampling 195

The replacement of an observation by a window is not standard in change-point

detection theory, since it requires using extra data points not typically available. While
the implementation may pose some difficulty in the environmental boundary-tracking
problem, there is no such problem for the corresponding image processing problem.
Besides noise, there can be other difficulties with the boundary tracker. Occa-
sionally, the boundary tracker becomes stuck in a certain area due to ambiguities in the
object boundaries or irregularity of its shape. To prevent this occurrence, the boundary

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

tracker receives a “kick” if it has not left a certain window by a certain number of iter-
ations. The kick is in the opposite direction to which the boundary tracker entered the
window.
More explicitly, starting at a point xi , let M1 be a constant representing the “win-
dow” size that the tracker must leave, and M2 a constant representing the number of
iterations before receiving a “kick.” If xi − xi+ j 2 < M1 for each j ≤ M2 , then the tracker
location xi+M2 is moved to xi+M2 − 2(
xi+M2 − xi ), and the new starting point for the kicking
algorithm becomes xi+M2 .
Theoretically, this direction is justified by assuming that the boundary tracker
enters the kicking window due to chance noise effects and that a slightly different path
can avoid problems. In experiments, the efficacy of kicking is confirmed in that after
going off course, the tracker often finds its way back to the boundary after kicking. In the
numerical experiments following, the tracker receives a kick in the opposite direction if
it has stayed in the same 5 × 5 window for 200 iterations.
Another improvement is to use the theoretical average detection delay to back-
track along the path. In cases where the structure of noise is known or can be estimated,
the average detection delay can often be calculated. Since this is the average time to
detection after a change-point has actually occurred, backtracking along the path by
this amount will give the correct change-point on average. This results in a more accu-
rate detection of the actual boundary points.
For the grayscale numerical experiments, average detection delay was approxi-
mated by the asymptotic detection delay, Ū
K(g, f)
. For f generated by N(0, σ 2 ), g generated
by N(1, σ 2 ),

g
K(g, f) = E g log
f
∞
g(y)
= g(y) log dy
−∞ f(y)
∞  2 
y (y − 1)2
= 2
− g(y) dy
−∞ 2σ 2σ 2
196 A. Chen et al.

∞  
(x + 1)2 x2
= − f(x) dx
−∞ 2σ 2 2σ 2
1
= .
2σ 2
Thus, the asymptotic detection delay is Ū ∗ 2σ 2 . Backtracking by Ū ∗ 2σ 2 steps
after each boundary crossing gives a slightly sharper tracking. The quantity Ū ∗ 2σ 2 is
normally small enough (< 1) that there is little difference in the actual results. For high

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

noise images, however, this backtracking has larger influence since both σ and Ū will
be larger.
The statistical nature of the CUSUM algorithm can be used to detect inaccurate
trackings. The idea is to observe whether the algorithm is making accurate tracking or
detecting an object only because it is forced between the two classes. If the object is not
being tracked accurately, the algorithm should switch from the local sampling algorithm
back to the global searching step to find a boundary point. Once a new boundary point
is found, local sampling can begin again.
Recalling the log-likelihood ratio Z k for each point along the tracking path,
N
tracking the statistic C = 1
N k=1 |Z k|, where N is the total number of points in the path,
gives a measurement of the confidence that the tracking is valid. If C > T̄ for some con-
stant threshold T̄, then one can be reasonably sure that a large portion of data points
are well-separated by the density functions f and g. In contrast for C < T̄, the track-
ing is likely to be inaccurate as boundary points are classified mainly because they are
forced into one class or the other. This is particularly a problem if the assumption of
a two-region decomposition for Ω is false. The threshold T̄ is determined empirically,
and a value of T̄ = 0.3 typically works well.

5 Numerical and Image Examples

In Figure 3, several results on a “U” image in heavy noise are shown. The results for
boundary tracking are comparable to a region-based method. Furthermore, corners are
not rounded as in the case with many energy-based segmentation methods. This fact
can make boundary tracking more well-suited for segmentation problems with more
man-made objects or with sharper edges.
Parameters are determined experimentally, but they are relatively consistent
from image to image. The primary adjustments are based on the level of noise present
in the image. For higher levels of noise, the CUSUM threshold should be set higher in
order to avoid false alarms, but this will result in higher detection delays. Visually,
one would see a larger sinusoidal pattern (a less sharp segmentation and an increasing
 Efficient Boundary Tracking Through Sampling 197

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Fig. 3. (3.1) A 100 × 100 image was corrupted with additive Gaussian noise, N(0,0.5). (3.2) Bound-
ary tracking without change-detection. (3.3) Boundary tracking with CUSUM. (3.4) Boundary
tracking with CUSUM and a 3 × 3 window average. (3.5) Threshold dynamics [23]. (3.6). Bound-
ary tracking on a 1000 × 1000 version of the “U” image.

chance of straying too far from the boundary), so it is best to avoid an unnecessarily
large threshold. The image resolution also affects the choice of parameters, but only
slightly. In general, a higher resolution gives more leeway for detection delays, since
the tracker will not have strayed far from the boundary. Thus, a higher threshold is
allowable. However, the algorithm works best for higher resolution images, and there is
considerable flexibility in choosing parameters for this case. Parameters for each of the
numerical examples were ω = 0.5, V = 0.5, Ū = L̄ = 0.8 unless otherwise indicated.
The segmentation accuracy of the boundary-tracking algorithm is comparable to
that of other segmentation methods. The run-time and storage costs, however, are much
less for boundary tracking than for many global methods, especially for high dimen-
sional data.
The algorithm compares favorably with other segmentation methods designed
for fast computation time. The split Bregman globally convex segmentation (GCS)
method of Goldstein and Osher [26] is one such method. Table 1 shows a comparison
of the two methods. Note that the boundary-tracking method scales much better with
an increase in image size.
198 A. Chen et al.

Table 1 A comparison of run-times for various images and segmentation methods. Computations
were done with an Intel Core 2 Duo Processor T8300 (2.4 GHz)
A comparison of run-times (s)

Image Boundary tracking Split Breg. GCS

100 × 100 U 0.0025 0.034
300 × 300 U 0.0089 0.620
1000 × 1000 U 0.0420 6.600

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

For the tracking of more than one object, it is possible to use another segmen-
tation method first on a subsampled version of the image to obtain topology and the
rough locations of objects. This has the added benefit of giving estimates for the mean
and variance, which can be used in the case of Gaussian noise. Furthermore, if the ini-
tial segmentation is reasonably accurate, then the boundary-tracking algorithm can be
restricted to detecting only boundary points close to those detected in the initial seg-
mentation. Thus, the boundary-tracking algorithm acts as a fast refinement to another
segmentation method. The resulting detection can also be used as an initialization to
another segmentation method on the full data set. The last modification would give fast
convergence while preserving the strengths of an alternative method.
An example on a simple, noisy image is shown in Figure 4. The original image
is 1000 × 1000. Threshold dynamics was first applied to a heavily subsampled version
(100 × 100) of the image. Then one pixel from each connected component was taken as
the starting point for a boundary tracker.
The boundary-tracking algorithm is ideally suited for segmentation in large data
sets. Detection delays of boundary points are the same in terms of number of steps,
whether for low resolution or high-resolution data. Thus, relative to the size of objects,
detection delays are much shorter. Furthermore, steps with smaller angular increment
can be used, effectively taking straighter steps and tracking the boundary more sharply.
Using a lower angular increment also increases the speed of tracking, lowering the num-
ber of iterations needed to track the object completely. A tracking of a 1000 × 1000 ver-
sion of the “U” image is also shown in Figure 3. Note that the “unit” of detection delay
is in steps. That is, for a high-resolution image, the detection delay is much smaller
in terms of the characteristic width of features. Thus, the tracking result looks much
more accurate. While the jagged tracking of the 100 × 100 version still exists in the
1000 × 1000 version, the scale of the jagged behavior is small compared with the fea-
tures and thus is almost invisible. Detection delay is affected by the threshold Ū , which
 Efficient Boundary Tracking Through Sampling 199

(4.1) (4.2)

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Fig. 4. A hybrid threshold dynamics—boundary-tracking segmentation on a 1000 × 1000 image.
(4.1) Initial segmentation by threshold dynamics. The image is subsampled by a factor of 10 on
each axis. (4.2) Final segmentation by boundary tracking, with starting points for the trackers
from the initial segmentation.

Fig. 5. Boundary tracking of the San Francisco Bay coastline [54]. In this example, multiple
trackers need to be used since the coastline boundary is not connected.

is set according to the noise level. There is thus a trade-off between spatial resolution
and the allowable noise level. Note that for global segmentation methods, higher image
resolution may give little benefit due to the increasingly high computation cost.
In the “San Francisco Bay” data set [54], a multispectral data set taken by the
Landsat 7 satellite, the datum is 3000 × 3500 pixels, with boundaries of the object of
interest touching the edge of the image.The normalized difference vegetation index
(NDVI), commonly used for water detection [52], is taken as the decision function. This
index is a combination of two images at the same spatial coordinates, taken at a red
wavelength band and an infrared wavelength band (Figure 5).
200 A. Chen et al.

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Fig. 6. Building segmentation of the “Urban” data set [63] with a spectral angle decision function.

The boundary-tracking method also has a natural extension to hyperspectral

data, which consists of a group of spatially co-registered images taken at different
wavelengths in the electromagnetic spectrum. Thus, each spatial pixel now has a vector-
valued “spectral signature” instead of a scalar intensity. In fact, it has been observed
that different objects have distinct spectral signatures, so that objects are readily dis-
tinguishable by observing spectral signatures.
In the tracking algorithm, all that is required is a decision function that deter-
mines whether the tracker is in one of two regions. Then replacing a simple threshold
by a suitable distance can still give a good boundary detection. One distance that has
been found to work especially well with hyperspectral data is the spectral angle dis-
tance [55, 67]:
u· v
Spectral angle = cos−1 . (7)
u v
If there are multiple objects to be detected, one may wish to compare two classes
of objects, rather than two specific objects. In this case, the decision function should
compare distance to one class (the minimum distance to an object in the class) with
distance to the other class. This comparison is especially useful in the case of hyper-
spectral imagery. The theoretical background on using such class comparisons is given
in [58]. The segmentation resulting from a class comparison between building references
and background references for the Urban data set, a hyperspectral data [63] set taken
by HYDICE, an airborne sensor, is shown in Figure 6.
 Efficient Boundary Tracking Through Sampling 201

With the change to a different decision function, the choice of f and g becomes
more unclear. It is difficult or impossible to estimate the structure of the data in terms of
spectral angle distance. By choosing a large number of sample points, it may, however,
be possible to create an approximation of f and g from the data itself.
With hyperspectral imagery, the number of spectral bands can number in the
hundreds. With this amount of data, it is often necessary to reduce the size of the data,
either in the spectral dimension or through spatial subsampling. Using the boundary-

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

tracking algorithm equipped with the spectral angle distance, segmentation can be fast
and accurate.
The Smith Island H20000508 southendBIP data set [3–6] is a large hyperspec-
tral data set obtained by an airborne hyperspectral scanner (HyMAP) on May 8, 2000,
over Smith Island, VA, a barrier island in the Virginia Coast Reserve. The boundary
between land and sea is often ambiguous, causing problems especially when the full
spectral information is not used. The NDVI of the data generally gives a good threshold
for land and sea, but sandy and swampy areas occupy a range of values that consti-
tute the ambiguous region. Considering spectral signatures, however, gives a clearer
indication of the various types of terrain. Moreover, it is also possible to assign these
ambiguous areas to either region according to user preference, simply by changing the
reference points for each region.
In Figure 7, each pixel is unmixed as a linear combination of various reference
spectral signatures using the L 1 unmixing model of [29]. The reference signatures used
are that of beach, dark dirt, light dirt, vegetation, light water, and dark water. Each pixel
is assigned to either a land class or water class based on the maximum abundance of the
material at that location. The plots of the reference signatures are shown in Figure 8.
Figure 9 indicates that tracking the maximum abundance shows the boundary more
clearly than using the NDVI result.
More generally, boundary tracking can be used to efficiently track features that
can be derived from the solution of an inverse problem. Since boundary tracking is a
subsampling algorithm, the number of pixels that require intense processing is mini-
mal. In certain cases, the decision function for each pixel may be a complex calculation
requiring a high run-time. This is especially true when calculating the decision func-
tion accurately is more computationally expensive as the boundary is approximated
with greater precision. The next example for fractals is an example of this fact.
Tracking fractal boundaries such as coastlines is another interesting example of
the efficiency of boundary tracking (see Figure 5). For fractals such as the Mandelbrot
set, the escape algorithm [21] can be used to calculate whether a point is in the set or
202 A. Chen et al.

(7.1) (7.2)

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Fig. 7. The Smith Island data set [3–6]. (7.1) Reference points used in the land/water classification
by full unmixing. The reference signatures are that of 1, dark dirt; 2, light dirt; 3, vegetation; 4,
beach; 5, dark water; 6, light water. (7.2) The land/water classmap using maximum abundance of
the unmixed result.

Fig. 8. (8.1) Plots of the reflectances of the reference signatures of the land class: dark dirt, light
dirt, vegetation, beach. (8.2) Plots of the reflectances of the reference signatures of the water class:
dark water, light water.

outside. While it can be computationally expensive to implement the algorithm for very
fine detail, tracking only the boundary can be much faster. In addition, the potential
detail is unlimited since the fractal is defined in continuous space rather than the dis-
crete space of images. Moreover, boundary tracking follows rough boundaries without
introducing extra smoothing, as in many segmentation methods with a penalty on the
 Efficient Boundary Tracking Through Sampling 203

(9.1) (9.2)

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Fig. 9. Tracking the boundary of the “Smith Island” hyperspectral data set. (9.1) Grayscale
boundary tracking using the NDVI as an indicator between land and sea. (9.2) Boundary track-
ing using unmixing on each pixel of the data set. Only pixels visited by the boundary tracker are
unmixed.

Fig. 10. Boundary tracking of the Mandelbrot set at 400 × 400 resolution. A higher magnification
of the boxed region is shown in Figure 11.

length of the boundary. This fact makes boundary tracking a method well suited for
dealing with fractal-like structure (Figure 10). Figure 11 shows various magnifications
of boundary tracking on the Mandelbrot set. Table 2 gives a timing comparison between
the boundary-tracking algorithm and the calculation of the escape algorithm at the cor-
responding resolution. In the results shown in Figure 11 and Table 2, the computations
on the escape algorithm for boundary tracking were done only at locations that the
204 A. Chen et al.

(11.1)

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

(11.2)

(11.3)

Fig. 11. Tracking the boundary of the Mandelbrot set at increasing levels of magnification with
the boxed region showing the domain for the next highest level of magnification. The resolution
of each image is 400 × 400. (11.1) 10 times magnification, 102 times magnification. (11.2) 103 times
magnification, 104 times magnification. (11.3) 105 times magnification, 106 times magnification.
 Efficient Boundary Tracking Through Sampling 205

Table 2 A table comparing the run-time of the boundary-tracking algorithm (BT) and calculation
of the full domain (FD) for the Mandelbrot set at the corresponding resolution (pixel width). The
precision of the escape algorithm (Escape) was adjusted according to the resolution, as needed
Mandelbrot set timing comparison

Pixel width Escape BT (s) Full FD (s)

7.51 × 10−3 400 0.7 5.8
7.51 × 10−4 103 1.2 21.5

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

7.51 × 10−5 4 × 103 7.8 152.6
7.51 × 10−6 104 32.6 529.7
7.51 × 10−7 4 × 104 145.9 2100.4
7.51 × 10−8 105 218.6 6078.6
7.51 × 10−9 4 × 105 861.5 22042.3

tracker visits, not on the entire 400 × 400 images. The tracking results are shown over-
laid on the full images for clarity.
Note that as the magnification is increased, the number of iterations in the
escape algorithm needed to resolve the boundary accurately also increases. This is due
to the fact that the points being examined are much closer to the actual boundary of the
Mandelbrot set, so that the escape algorithm for points not in the Mandelbrot set will
diverge much more slowly. This gives an additional savings in the run-time of boundary
tracking relative to computing on the entire domain, as the resolution is increased.
By tracking the boundary of the fractal at various magnification factors, it is
possible to measure the fractal dimension using the box-counting method. The fractal
dimension of a set is defined to be
log(N())
D = lim ,
→0 log(1/)

where N represents the number of boxes of size  that are needed to cover the set. At a
given scale , there exists a simple relationship between the length L() and the number
of boxes N(): L() =  N(). Using this formula, the fractal dimension can be rewritten as

log(L/) log(L)
lim = 1 − lim .
→0 log(1/) →0 log()

Using various step sizes V for the scale parameter , boundary tracking yields
boundary points at various scales. Thus, a log–log plot of log(L) versus log(V) is approx-
imately a line with slope equal to one minus the fractal dimension. A calculation of the
boundary for the Julia set with c = 1
4
is shown in Table 3. The calculated slope is
206 A. Chen et al.

Table 3 A table of the length calculations for the Julia set at various scales
Julia set lengths

V Boundary points Length

5 × 10−5 42,582 3.7706
10−4 20,152 3.5642
3 × 10−4 6,210 3.2655
5 × 10−4 3,538 3.1364

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

10−3 1,668 2.9773
3 × 10−3 513 2.7254
5 × 10−3 308 2.5925
10−2 137 2.4744

Fig. 12. (12.1) The Julia set for c = 14 . (12.2) Log–log plot of log10 (L) versus log10 (V).

−0.0799, so the measured fractal dimension is 1.0799, which is close to the theoreti-
cal value of 1.0812 [41]. The corresponding set and log–log plot is shown in Figure 12.
Note that the Mandelbrot has comparatively a more complex structure, having a frac-
tal dimension of 2 [56], and the numerical computation of the fractal dimension is not
usually done. The computation of the fractal dimension here is a demonstration of one
application of boundary tracking.
Similar fractal dimension calculations can be done for real coastlines, as long
as the resolution of the data set is sufficient. For the Smith Island data set, a fractal
dimension can be calculated using boundary tracking. The results are shown in Table 4
 Efficient Boundary Tracking Through Sampling 207

Table 4 A table of the length calculations for the Smith Island set at various scales
Smith Island set lengths

V Boundary points Length

0.5 14,431 8,162
1 6,111 7,057
2 2,939 6,830
4 1,271 6,024

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

8 464 4,651
16 199 4,651
32 93 3,976

Fig. 13. Log–log plot of log10 (L) versus log10 (V) for Smith Island.

and Figure 13. From the fitted slope of −0.1739 for the log–log plot, the measured fractal
dimension is thus 1.1739.

6 Discussion

The greatest advantage of the boundary-tracking algorithm over region-based methods

for segmentation is the computation speed. The particular combination of boundary
tracking and change-point detection provides a simple, but accurate and computation-
ally effective segmentation method. Calculations and measurements are taken over sam-
ples, with the number of samples having the same order as the boundary length of
the tracked object. In region-based methods, however, calculations and measurements
are done over the entire image. Thus, boundary tracking is O(n), while region-based
208 A. Chen et al.

methods are O(n2 ), where the size of the image is n× n. Then the savings becomes more
apparent for larger images and those with higher resolution.
There is a substantial savings in memory requirements as well. Since only local
information is used in the decision function, it is not even necessary for the entire image
to be stored at once. Instead, image values can be read as needed whenever the tracker
visits a new point. On the other hand, many segmentation methods require not only the
original data to be stored but also evolved data of the same size.

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

With high-noise data, change-point detection is vital in minimizing mistakes
due to noise. Parameters are often chosen in order to minimize detection delay and
false alarms. In the boundary-tracking problem, the former is of paramount impor-
tance, since region changes happen very frequently. While false alarms can be tolerated
to some extent, consistent detection delays can seriously hamper the algorithm’s per-
formance, causing false alarms to increase as well, by leading the tracker away from
the boundary. Improving the algorithm’s ability to recover from such mistakes is an
important consideration.
The primary case in which the proposed algorithm fails is for high noise (or
textured), low resolution images, especially when a reasonable estimate of the charac-
teristics of each region is also unavailable. In high noise, boundary tracking may fail
with certain realizations of noise while still succeeding in others. It may be of interest
to explore failure rates and whether further modifications such as a better global search
procedure may make boundary tracking more robust to noise.
The change-point modifications run very quickly, requiring only a minimal
amount of extra storage and computation over the boundary-tracking algorithm alone.
Using a window replacement is slightly slower since more pixels are sampled. In this
case, however, the algorithm still operates in O(n). Moreover, the greater accuracy with
which the boundary is tracked can lead to additional savings in computation time, as
fewer tracking mistakes are made and thus fewer iterations are needed.
While it is desirable in general to minimize the false alarms and detection delay,
these indices do not provide a complete understanding of a change detection algorithm’s
effectiveness in the boundary-tracking problem. This is due to the fact that a false alarm
of just one change point can cause the boundary tracker to travel far from the bound-
ary. In subsequent readings, change-point detection may be completely accurate and
have low detection delay, but give no useful information about the boundary. Thus, the
effectiveness of a change point detection algorithm must be measured visually with
respect to the final segmentation result as well. Nevertheless, these parameters are
useful numerical indicators of a procedure’s effectiveness.
 Efficient Boundary Tracking Through Sampling 209

The change-point detection literature studies various measures of false alarms

and detection delays. Due to the interdependency of each change-point detection test
that is run, simply tracking these two indices may not be a good indicator of how the
algorithm performs. Also, most of the examples presented involve some a priori assump-
tion on the probability distribution of the data. This assumption may not be accurate
for some data, especially for complicated data sets like hyperspectral imagery. Some
possibilities from change-point detection [9] have been studied, but it is also not clear

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

whether these would represent an improvement over a priori assumptions on the distri-
butions.
Another important factor in the effectiveness of the algorithm is the character-
istic width of the object to be tracked. If the step size of the boundary tracker is of the
same order of magnitude as the characteristic width of the object, it can be difficult
to track an object effectively. But with a characteristic width much larger than bound-
ary tracker step size (equivalently, higher resolution) data, the algorithm can be highly
effective without a corresponding large jump in run-time. Note that step size cannot be
decreased arbitrarily; the resolution restricts the amount of data that can be used when
the step size becomes too small. Thus, narrow objects are difficult to track if the resolu-
tion of the data is not sufficient. Of course, this limitation on tracking features smaller
than characteristic dimensions is true for all segmentation methods.
When there are more than two regions to be differentiated, different possibilities
arise. One method is a modification to track classes of objects. This is especially relevant
to the case with hyperspectral data, in which there are multiple objects with different
spectral signatures. In this case, the decision function can be altered so that a given
pixel is measured against proximity to an entire class, rather than just to one object.
Particular care needs to be taken, however, to choose the classes carefully.
An extension to the boundary-tracking algorithm is to use multiple trackers.
The algorithm is easily parallelizable, adding a stopping condition for each tracker; for
example, the algorithm for one tracker terminates if it meets the path taken by any other
tracker. Aside from parallelization for the sake of run-time, information from multiple
trackers can potentially be used in concert to improve tracking.
The boundary-tracking algorithm can be of use in many problems in which run-
time is important. For video tracking, the large number of frames used can lead to
a large computation time. By using the boundary-tracking algorithm with initializa-
tion taken from previous frames, objects can be found and tracked quickly through
each frame. Other applications include surface segmentation and more sophisticated
parameter estimation models by using shape priors.
210 A. Chen et al.

Acknowledgments

The authors would like to thank Dr C. Bachmann and the Naval Research Laboratory
for helpful discussions on geospatial tracking problems as well as the use of the “Smith
Island” data set, Z. Hu and V. Mejia for earlier work on the boundary-tracking algorithm
in the image processing context [32, 42], and T. Goldstein for the implementation of
the split Bregman GCS method. The authors would also like to thank the reviewers

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

for numerous comments that have improved the presentation and organization of the
manuscript.

Funding

This work was supported by Office of Naval Research grant N000140810363; the Depart-
ment of Defense; ARO MURI grant 50363-MA-MUR; and National Science Foundation
grant DMS-0914856.

References
[1] Andersson, S. “Curve tracking for rapid imaging in AFM.” IEEE Transactions on Nanobio-
science 6, no. 4 (2007): 354–61.
[2] Azzabou, N., N. Paragios, and F. Guichard. “Random walks, constrained multiple hypotheses
testing and image enhancement.” In 9th European Conference in Computer Vision (ECCV),
379–90. Graz, Austria, May 7–13, 2006.
[3] Bachmann, C. “Improving the performance of classifiers in high-dimensional remote sensing
applications: An adaptive resampling strategy for error-prone exemplars (ARESEPE).” IEEE
Transactions on Geoscience and Remote Sensing 41, no. 9 (2003): 2101–12.
[4] Bachmann, C., T. Ainsworth, and R. Fusina. “Exploiting manifold geometry in hyperspectral
imagery.” IEEE Transactions on Geoscience and Remote Sensing 43, no. 3 (2005): 441–54.
[5] Bachmann, C., T. Ainsworth, and R. Fusina. “Improved manifold coordinate representations
of large scale hyperspectral imagery.” IEEE Transactions on Geoscience and Remote Sensing
44, no. 10 (2006): 2786–803.
[6] Bachmann, C., T. Donato, G. Lamela, W. Rhea, M. Bettenhausen, R. Fusina, K. Du Bois,
J. Porter, and B. Truitt. “Automatic classification of land cover on Smith Island, VA, using
HyMAP imagery.” IEEE Transactions on Geoscience and Remote Sensing 40, no. 10 (2002):
2313–30.
[7] Bai, X. and G. Sapiro. “Distancecut: Interactive segmentation and matting of images and
videos.” In IEEE ICIP 2, 249–52, San Diego, CA, October 12–15, 2007.
[8] Barchiesi, M., S. Kang, T. Le, M. Morini, and M. Ponsiglione. “A variational model for infinite
perimeter segmentations based on Lipschitz level set functions: denoising while keeping
 Efficient Boundary Tracking Through Sampling 211

finely oscillatory boundaries.” SIAM Multiscale Modeling and Simulation 8, no. 5 (2010):
1715–41.
[9] Basseville, M. and I. Nikiforov. Detection of Abrupt Changes—Theory and Applications.
Information and System Sciences Series. Englewood Cliffs, NJ: Prentice Hall, 1993.
[10] Bertozzi, A. and J. Greer. “Low-curvature image simplifiers: Global regularity of smooth
solutions and Laplacian limiting schemes.” Communications on Pure and Applied
Mathematics 57, no. 6 (2004): 764–90.

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

[11] Canny, J. “A computational approach to edge detection.” IEEE Transactions on Pattern
Analysis and Machine Intelligence 8, no. 6 (1986): 679–98.
[12] Castleman, K. Digital Image Processing. Englewood Cliffs, NJ: Prentice Hall, 1996.
[13] Chan, T. and L. Vese. “Active contours without edges.” IEEE Transactions on Image
Processing 10, no. 2 (2001): 266–77.
[14] Chang, P. and S. Andersson. “Smooth trajectories for imaging string-like samples in AFM: A
preliminary study.” In 2008 American Control Conference, Seattle, WA, June, 2008.
[15] Chen, A., T. Wittman, A. Tartakovsky, and A. Bertozzi. “Image segmentation through effi-
cient boundary sampling.” In Proceedings of the 2009 Workshop on Sampling Theory and
Applications, Marseille, France, May, 2009.
[16] Chiang, C. and J. Liu. “Boundary following in unknown polygonal environment based on fast
marching method.” In IEEE International Conference on Advanced Robotics and its Social
Impacts, Taipei, Taiwan, August 23–25, 2008.
[17] Cohen, L. “On active contour models and balloons.” CVGI: Image Understanding 53, no. 2
(1991): 211–8.
[18] Couzin, I. and N. Franks. “Self-organized lane formation and optimized traffic flow in army
ants.” Proceedings of the Royal Society of London B Biology 270 (2003): 139–46.
[19] Darbon, J. A Note on the Discrete Binary Mumford–Shah Model, 283–294. Lecture Notes in
Computer Science 4418. Berlin: Springer, 2007.
[20] del Castillo, E. Statistical Process Adjustment for Quality Control. Wiley Series in Probabil-
ity and Statistics. New York: Wiley, 2002.
[21] Douady, A. and J. Hubbard, with P. Lavaurs, T. Lei and P. Sentenac. “Etude dynamique des
polynômes complexes.” Société Mathématique de France, Documents Mathématiques. Paris,
France, 2007.
[22] El-Zehiry, N., S. Xu, P. Sahoo, and A. Elmaghraby. “Graph cut optimization for the Mumford-
Shah model.” In Proceedings of the Seventh IASTED International Conference Visualization,
Imaging, and Image Processing, Palma de Mallorca, Spain, August 2007.
[23] Esedoglu, S. and Y. Tsai. “Threshold dynamics for the piecewise constant Mumford-Shah
functional.” Journal of Computational Physics 211, no. 1 (2006): 367–84.
[24] Geman, S. and D. Geman. “Stochastic relaxation, Gibbs distributions, and the Bayesian
restoration of images.” IEEE Transactions on Pattern Analysis and Machine Intelligence
6, no. 6 (1984): 721–41.
[25] Geman, D. and B. Jedynak. “An active testing model for tracking roads in satellite images.”
IEEE Transactions on Pattern Analysis and Machine Intelligence 18, no. 1 (1996): 1–14.
212 A. Chen et al.

[26] Goldstein, T. and S. Osher. “The split Bregman algorithm for L1 regularized problems.” SIAM
Journal on Imaging Sciences 2, no. 2 (2009): 323–43.
[27] Gonzalez, R. and R. Woods. Digital Image Processing. Englewood Cliffs, NJ: Prentice Hall,
2008.
[28] Griffiths, T. and A. Yuille. “Technical introduction: A primer on probabilistic inference.”
January, 2006.
[29] Guo, Z., T. Wittman, and S. Osher. “L1 unmixing and its application to hyperspectral image

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

enhancement.” UCLA CAM report, March, 2009.
[30] Hangartner. W. “Spezifität und inaktivierung des spurpheromons von lasius fuliginosus
(latr.) und orientierung der arbeiterinnen im duftfeld.” Zeitschrift für Vergleichende Phys-
iologie 57 (1967): 103–36.
[31] Hsieh, C., Z. Jin, D. Marthaler, B. Nguyen, D. Tung, A. Bertozzi, and R. Murray. “Experimen-
tal validation of an algorithm for cooperative boundary tracking.” In Proceedings of the
American Control Conference, 1078–83. Portland, OR, June 8–10, 2005.
[32] Hu, Z. “Gradient-free boundary tracking.” Undergraduate research project.
[33] Huang, Z. and K. Chau. “A new image thresholding method based on Gaussian mixture
model.” Applied Mathematics and Computation 205, no. 2 (2008): 899–907.
[34] Jin, Z. and A. Bertozzi. “Environmental boundary tracking and estimation using multiple
autonomous vehicles.” In 2007 46th IEEE Conference on Decision and Control, 4918–4923.
New Orleans, LA, December 2007.
[35] Joshi, A., T. Ashley, Y. Huang, and A. Bertozzi. “Experimental validation of cooperative envi-
ronmental boundary tracking with on-board sensors.” In American Control Conference,
2630–5, St. Louis, MO, June 10–12, 2009.
[36] Kass, M., A. Witkin, and D. Terzopoulos. “Snakes: Active contour models.” International
Journal of Computer Vision 1, no. 4 (1988): 321–31.
[37] Kemp, M., A. Bertozzi, and D. Marthaler. “Multi-UUV perimeter surveillance.” In Proceedings
of 2004 IEEE/OES Workshop on Autonomous Underwater Vehicles, 102–7, Sebasco Estates,
ME, June, 2004.
[38] Leung, K., C. Hsieh, Y. Huang, A. Joshi, V. Voroninski, and A. Bertozzi. “A second generation
micro-vehicle testbed for cooperative control and sensing strategies.” In Proceedings of the
2007 American Control Conference, 1900–7, New York, NY, July 11–13, 2007.
[39] Liu, W., M. Short, Y. Taima, and A. Bertozzi. “Multiscale collaborative searching through
swarming.” In Proceedings of the 7th International Conference on Informatics in Control,
Automation, and Robotics (ICINCO), Portugal, June 15–18, 2010.
[40] Lorden, G. “Procedures for reacting to a change in distribution.” Annals of Mathematical
Statistics 42, no. 6 (1971): 1897–908.
[41] McMullen, C. “Hausdorff dimension and conformal dynamics, III: Computation of
dimension.” American Journal of Mathematics 120, no. 4 (1998): 691–721.
[42] Mejia, V. “Multiple robot boundary tracking.” Undergraduate research project.
[43] Montgomery, D. Introduction to Statistical Quality Control. New York: Wiley,
1985.
 Efficient Boundary Tracking Through Sampling 213

[44] Moustakides. G. “Optimal stopping times for detecting changes in distributions.” Annals of
Statistics 14, no. 4 (1986): 1379–87.
[45] Mumford, D. and J. Shah. “Optimal approximation by piecewise smooth functions and asso-
ciated variational problems.” Communications on Pure and Applied Math XLII, no. 5 (1989):
577–684.
[46] Ng, J. and T. Bräun. “Performance comparison of bug navigation algorithms.” Journal of
Intelligent and Robotic Systems 50, no. 1 (2007): 73–84.

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

[47] Osher, S. and J. Sethian. “Fronts propagating with curvature-dependent speed: Algorithms
based on Hamilton-Jacobi formulation.” Journal of Computational Physics 79, no. 1 (1988):
12–49.
[48] Page, E. “Continuous inspection schemes.” Biometrika 41, no. 1–2 (1954): 100–115.
[49] Peng, D., B. Merriman, S. Osher, H. Zhao, and M. Kang. “A PDE-based fast local level set
method.” Journal of Computational Physics 155, no. 2 (1999): 410–38.
[50] Protiere, A. and G. Sapiro. “Interactive image segmentation via adaptive weighted distances.”
IEEE Transactions on Image Processing 16, no. 4 (2007): 1046–57.
[51] Rother, C., V. Kolmogorov, and A. Blake. “GrabCut: Interactive foreground extraction using
iterated graph cuts.” In ACM Transactions on Graphics (SIGGRAPH), August 2004.
[52] Rouse, J., R. Hass, J. Schell, and D. Deering. “Monitoring vegetation systems in the grain
plains with ERTS.” In Third ERTS Symposium, 309–317. NASA SP-351 I, 1973.
[53] Ryan, T. Statistical Methods for Quality Improvement, 2nd ed. Wiley Series in Probability
and Statistics. New York: Wiley, 2000.
[54] San Francisco Bay multispectral data set. Landsat 7 satellite images.
[55] Shippert, P. “Introduction to hyperspectral image analysis.” Online Journal of Space
Communication 3 (Winter 2003).
[56] Shishikura, M. “The Hausdorff dimension of the boundary of the Mandelbrot set and Julia
sets.” Annals of Mathematics 147, no. 2 (1998): 225–67.
[57] Stauffer, C. and W. Grimson. “Adaptive background mixture models for real-time tracking.”
In Proceedings of Computer Vision Pattern Recognition June 1999.
[58] Tartakovsky, A. “Asymptotic performance of a multichart CUSUM test under false alarm
probability constraint.” In Proceedings of the 44th IEEE Conference on Decision and Control,
and the European Control Conference, 320–5, Seville, Spain, December 12–15, 2005.
[59] Tartakovsky, A., B. Rozovskii, R. Blažek, and H. Kim. “Detection of intrusions in informa-
tion systems by sequential change-point methods.” Statistical Methodology 3, no. 3 (2006):
252–93.
[60] Tartakovsky, A. and V. Veeravalli. “Change-point detection in multichannel and dis-
tributed systems with applications.” In Applications of Sequential Methodologies, edited
by N. Mukhopadhyay, S. Datta, and S. Chattopadhyay, 331–363. New York: Marcel Dekker,
2004.
[61] Tu, Z. and S. Zhu. “Image segmentation by data-driven Markov chain Monte Carlo.”
IEEE Transactions on Pattern Analysis and Machine Intelligence 24, no. 5 (2002):
657–73.
214 A. Chen et al.

[62] Tumblin, J. and G. Turk. “LCIS: A boundary hierarchy for detail-preserving contrast
reduction.” In SIGGRAPH Annual Conference on Computer Graphics, 83–90, Los Angeles,
CA, August 8–13, 1999.
[63] Urban hyperspectral data set. HYDICE sensor imagery.
[64] Viswanathan, G., S. Buldyrev, S. Havlin, M. da Luz, E. Raposo, and E. Stanley. “Optimizing
the success of random searches.” Nature 401, no. 6756 (1999): 911–4.
[65] Willsky, A. “A survey of design methods for failure detection in dynamical systems.”

Downloaded from https://academic.oup.com/amrx/article/2011/2/182/102168 by guest on 25 November 2024

Automatica 12, no. 6 (1976): 601–11.
[66] Xu, C. and J. Prince. “Gradient vector flow: A new external force for snakes.” Proceedings
of theIEEE Conference on Computer Vision Pattern Recognition (CVPR), 66–71, Puerto Rico,
June 17–19, 1997.
[67] Yunas, R., A. Goetz, and J. Boardman. “Discrimination among semi-arid landscape endmem-
bers using the spectral angle mapper (SAM) algorithm.” In Summaries of the Third Annual
JPL Airborne Geoscience Workshop, 147–9, Pasadena, CA, June 1–5, 1992.
[68] Zhu, S. and A. Yuille. “Region competition: unifying snakes, region growing, and Bayes/MDL
for multiband image segmentation.” IEEE Transactions on Pattern Analysis and Machine
Intelligence 18, no. 9 (1996): 884–900.