Mean-shift Blob Tracking through Scale Space 

Robert T. Collins
Carnegie Mellon University

Abstract
The mean-shift algorithm is an efficient technique for tracking 2D blobs through an image. Although the scale of the
mean-shift kernel is a crucial parameter, there is presently
no clean mechanism for choosing or updating scale while
tracking blobs that are changing in size. We adapt Lindebergs theory of feature scale selection based on local maxima of differential scale-space filters to the problem of selecting kernel scale for mean-shift blob tracking. We show
that a difference of Gaussian (DOG) mean-shift kernel enables efficient tracking of blobs through scale space. Using
this kernel requires generalizing the mean-shift algorithm
to handle images that contain negative sample weights.

1. Introduction
The mean-shift algorithm is a nonparametric statistical
method for seeking the nearest mode of a point sample distribution [3, 6]. The algorithm has recently been adopted as
an efficient technique for appearance-based blob tracking
[1, 4]. In the blob tracking scenario, sample points are regularly distributed along the image pixel grid, with a pixels
value w(a) being the sample weight of the point at that location. This sample weight is chosen such that pixels on the
foreground blob have high weight, and pixels in the background have low weight. The mean-shift algorithm specifies how to combine the sample weights w(a) in a local
neightborhood with a set of kernel weights K (a) to produce
an offset that tracks the centroid of the blob in the image
(Figure 1).
The problem we address in this paper is selecting the
scale of the mean-shift kernel, which directly determines
the size of the window within which sample weights are examined. This size should of course be proportional to the
expected image area of the blob being tracked. Although
kernel scale is a crucial parameter for the mean-shift algorithm, there is currently no sound mechanism for choosing
this scale within the framework. We show how to combine a
well-developed theory of feature scale selection due to Lindeberg with the mean-shift algorithm, resulting in a method
 This work is supported in part by DARPA/IAO HumanID under
ONR contract N00014-00-1-0915, and by DARPA/IPTO MARS contract
NBCHC020090.

Figure 1: Two-dimensional blob tracking by applying the

mean-shift algorithm to an image where pixel values represent likelihood of being on the tracked object. The meanshift algorithm is a non-parametric method for climbing to
the nearest mode of this likelihood image.
that represents and tracks blobs as the modes of a differential function in scale space.
In this paper we do not specify how the sample weight
image w(a) is produced. Typically, weights are determined
using a color-based object appearance model. In [1] a histogram of skin color in HSV color space is used to determine the likelihood of skin occuring at each pixel, using histogram backprojection to replace each pixel with the
frequency (probability) associated with that HSV value in
the skin color histogram. In [4] the appearance model is a
distribution of colors represented by a histogram m, which
is compared with a histogram of colors d observed within
the current mean-shift window. The sample weight at each
pixel with color i is then set to mi =di , a value related to
the Bhattacharyya coefficient of histogram similarity. The
mean-shift tracking paradigm applies equally well to sample weight images computed using other features besides
color, such as texture similarity, background subtraction results, or the output of intensity template correlation.
In Section 2 we review the mean-shift algorithm and its
relationship to non-parametric mode seeking. We also discuss how to generalize the algorithm to handle images with
negative sample weights. In Section 3 we review Lindebergs theory of feature scale selection. This theory represents image features of different scales by local maxima of
a differential operator in scale-space. Section 4 combines

the two methods to show how mean-shift tracking of a blob

that changes in scale can be performed by tracking the blob
features mode through scale space. The paper closes by
showing an illustrative example.

2. Mean-shift and Its Limitations

a K (a x)w(a)a
a K (a x)w(a)

a K (a x)w(a)(a x)
a K (a x)w(a)
(1)
where K is a suitable kernel function and the summations
are performed over a local window of pixels a around the
current location x. A suitable kernel K is one that can
be written in terms of a profile function k such that K (y) =
k(kyk2 ) and profile k Ris nonnegative, nonincreasing, piecewise continuous, and 0 k(r)dr < [Cheng].
An important theoretical property of the mean-shift algorithm is that the local mean-shift offset x computed at
position x using kernel K points opposite to the gradient direction of the convolution surface
x=

C(x) = H (a

x)w(a)

=
=

The heart of the mean-shift algorithm is computation of an

offset from location vector x to a new location x0 = x + x ,
according to the mean-shift vector
x =

explored to date are symmetric about the origin, such that

K (x) = K ( x), in which case

(2)

computed with a kernel H, which is called the shadow of

kernel K. This property makes the mean-shift procedure
an efficient iterative method for gradient ascent to a local
mode of the convolution surface without having to compute
the full convolution.
Kernels K and shadow kernels H must satisfy the relationship
(3)
h0 (r) = ck(r)
where h and k are the respective profiles of H and K, r =
ka xk2, and c > 0 is some constant [3]. Equation 3 guarantees that the mean-shift vector x computed from equation 1
points in the opposite direction to the gradient C(x) of the
surface in equation 2 for all x. For example, consider the
Gaussian shadow kernel H (x) = expf ka xk2 =(22 )g,
which has profile h(r) = expf r=(22)g. Since h0 (r) =
1
expf r=(22)g we find from equation 3 that k(r) =
22
expf r=(22 )g, and thus the associated kernel is K (x) =
expf ka xk2 =(22 )g, showing that the Gaussian kernel
is its own shadow. Another popular kernel-shadow pair
is the flat kernel fK (x) = 1; kxk < 1g and its associated
shadow fH (x) = 1 kxk2; kxk < 1g.
Note that the mean-shift vector computed by equation 1
is invariant to scaling of the sample weights w(a) by a positive constant c, that is, if each w(a) is replaced by cw(a)
for c > 0. Note that the mean-shift vector is not invariant to
a constant offset w(a) + c. However, all mean-shift kernels

K (a x)(w(a) + c)(a x)
K (a x)(w(a) + c)
K
(a x)w(a)(a x) + c K (a x)(a
K (a x)(w(a) + c)
K (a x)w(a)(a x)
K (a x)(w(a) + c)

(4)
x)

(5)
(6)

showing that the direction of the mean-shift vector is invariant, and just the step size changes.

2.1 Mean-shift and negative weights

One limitation of the classic mean-shift algorithm is that the
sample weights w(a) have to be nonnegative. Indeed, empirical observation of mean-shift performance in the presence of negative sample weights shows that the method
breaks down, with the algorithm sometimes seeming to converge to almost anything except the blob being tracked. To
explain this behavior, consider what happens when c in
equation 6 is a negative number. Although the direction of
the mean-shift offset (numerator) is still valid, if c is larger
than some of the values w(a), the step size (denominator)
can become negative, and the algorithm takes a step away
from the mode. Once this begins to happen, the algorithm
quickly diverges.
Why might we want to use negative sample weights?
When comparing similarity of a model color histogram m
with data histogram d, an intuitive alternative to mi =di
would be the log likelihood value log(mi =di ), which is negative when di > mi . Accumulating log likelihood values
with a flat kernel would then amount to computing the KLdivergence between the histograms m and d, that is

mi

log di
a

=n

mi

di log di

i=1

A second reason for wanting to use negative values is that

we might try to make the mean-shift vector invariant to both
scale AND offset of the sample weights w(a). We can easily
do this by subtracting the sample mean w from the values
w(a) within the kernel window, thereby performing meanshift on the offset-normalized image w(a) w. A third reason for allowing negative weights is that we might want to
use mean-shift kernels that have negative weights. This last
point is explored in Section 4.
Before changing the mean-shift algorithm to handle negative sample weights, we must first determine in what way
negative weights make sense. If we interpret the mean shift
quantity (Eq 1) as a sample center of mass, then a negative mass at any point doesnt make any sense. Instead,

we should interpret a points weight as a vote for the direction and magnitude of the mean shift offset vector towards or away from that point. For a single point p, a positive weight w specifies that the offset vector at neighboring
points should be directed towards p, with a magnitude w.
If instead we have a negative weight w, the offset vector at
neighboring points is now directed away from p, but with
magnitude jwj. The mean-shift equation can be interpreted
as forming the superposition of all these point-wise offset
votes to produce an overall average offset vector.
We now see how to modify the mean-shift equation (1)
to make sense for negative weights. The numerator of
that equation votes for both the magnitude and direction
of point-wise offset vectors, so the negative weights should
stay. However, the denominator normalizes by the overall
total magnitude of the votes, and therefore we must sum
only the magnitude (the absolute value) of each term. The
modified equation is
x =

a K (a x)w(a)(a x)
a jK (a x)w(a)j

(7)

With this modification, mean-shift remains well-behaved on

images that contain negative pixel values.

2.2 Mean-shift and scale selection

Kernel scale is a crucial parameter to the performance of the
mean-shift algorithm. If the kernel size is chosen too large,
the tracking window will contain many background pixels
as well as the foreground object pixels. This is a problem
when a data histogram collected within the window is compared to a model histogram describing the appearance of
the foreground object, since the data histogram is then diluted with noisy background pixels. A window size that is
too large can also cause the tracker to become more easily
distracted by background clutter. Finally, a kernel window
that is too large may cause convergence to an area between
multiple modes, rather than converging to just one of the
modes (see Figure 2, left).
Choosing a kernel size that is too small is also a problem. Kernels that are too small can roam around on a
likelihood plateau around the mode, leading to poor object
localization (Figure 2, right). This is particularly bad when
mean-shift tracking results are being used to point an active
pan/tilt camera. In this case, even when the object is not
moving, the pan/tilt head will randomly jump around as the
mean-shift window moves around the likelihood plateau.
The is no natural mechanism within the mean-shift
framework for choosing the kernel scale or adapting it over
time. [1] uses moments of the sample weight image to compute blob scale and orientation. This is only appropriate if
there is a single foreground object in the scene. [4] suggests
repeating the mean-shift algorithm at each iteration using

Figure 2: Mean-shift kernels that are too big (left) include

background clutter as well as the foreground object pixels,
and can also fail by encompassing multiple modes. Kernels
that are too small (right) can roam around on a likelihood
plateau around the mode, leading to poor localization.
window sizes of plus or minus 10 percent of the current size,
and evaluating which scale is best using the Bhattacharyya
coefficient. We have experimented with this approach and
found it lacking. In particular, the approach does keep the
window from growing too large, but unfortunately does not
stop the window from shrinking too small. Looking at the
right hand side of Figure 2, we see that in a uniformly colored region, any location of a window that is too small will
yield the same value of the Bhattacharyya coefficient, as
will shrinking the window even more. In short, this method
has no expansion force to keep the window inflated. In
[5], methods for adaptive kernel size selection are explored
in the context of clustering point sample distributions for
image segmentation, however we believe these methods are
currently too computationally expensive for real-time tracking applications.

3. Scale Space Blobs

In this section we consider a method for automatically
choosing an appropriate scale for blob tracking. The work
of Lindeberg provides an elegant theory for selecting the
best scale for describing features in an image [8, 2]. A
multi-scale image description is formed (at least conceptually) by convolving the image with Gaussian filters of increasing spatial variance to form a scale-space representation with two spatial dimensions and a third dimension representing scale or resolution. Lindeberg then proposes to
detect image features as points in scale-space where various
differential operators achieve a local maxima with respect to
both space and scale. This approach has been shown to produce feature descriptions that are well-localized both spatially and at an appropriate level of resolution.
For blob-like features, the differential operator used is
the Laplacian, which when applied over Gaussian scalespace leads to a multi-resolution image description based
on convolution with Laplacian-of-Gaussian (LOG) filters at
varying scales. The 2D LOG filter at scale has the form
LOG(x; ) =

22 kxk2 kx2k2
e 2
26

(8)

Let the convolution of image f (x) with a LOG operator of

scale be denoted as L(x; ) = a LOG(a x; ) f (a) :
Then blob features at various scales can be detected as
points (x0 ; 0 ) that are local maxima both spatially and in
scale, i.e.
(x

L)(x0 ; 0 ) = 0 and ( L)(x0 ; 0 ) = 0

(9)

Let image f (x) contain a blob with scale , centered at

x0 , so that L(x0 ; ) is a local maximum in the scale space
representation of f . Consider a new image g(y) = f (y=s)
that is a spatially scaled version of f . As we would then
expect, L(sx0 ; s) is a local maxima in the scale space representation of g. However, the magnitudes of the two filter
responses are different, which makes it hard to compare feature decriptions across different scales. Assuming without
loss of generality that the blob is centered at zero, and performing a change of variables y = sx on L(0; ) we find that
L(0; )

22 kxk2 kx2k2
e 2 f (x)
26
22

s2

ky sk2
=

22

f (y=s) j

1
j
s2

(12)

s2 L(0; s)

(13)

22 =1:6

kxk2

e 22

16

= :

1
22 (1:6)

 0 4875 2 LOG(x; )
:

The extra 2 factor is what makes the DOG operator invariant across scales. This precise manner in which the DOG
operator approximates the LOG operator seems not to be
widely discussed. The two are typically presented in the
context of locating zero crossings, where the scale factor
does not matter.

4. Mean-shift in Scale Space

kyk2

where the j1=s2 j term at the end of Eq. 11 is the Jacobian

of the transformation x = y=s. In general, Lindeberg shows
that L(x; ) = s2 L(sx; s), and defines a new normalized operator 2 LOG(x; ) whose magnitude stays invariant across
changes of image scale [8].
In the present work, we propose to replace the LOG scale
space with one based on Difference-of-Gaussian (DOG) filters. It is well-known that the LOG filter can be wellapproximated with a difference of two centered Gaussians
G(x; 0; 1 ) G(x; 0; 2 ) with scales related by 2 =1 = 1:6
[7]. The DOG filter that approximates the LOG filter at
scale thus has the form
DOG(x; ) =

DOG(x; )

(11)

kyk e 2(s)2 g(y)

2(s)6

2(s)2

26

ky sk2 e

(10)

with a DOG operator is written as D(x; ) = a DOG(a

x; ) f (a) ; then performing a change of variables procedure
as above shows that D(x; ) = D(sx; s), so we do not have
to normalize the DOG operator to achieve invariance across
scales.
Why does the LOG filter have to be normalized to give
an invariant response across scales, while its approximation,
the DOG filter, does not? The DOG operator does not approximate the LOG in the traditional sense of DOG(x) 
LOG(x). Rather, the two functions are approximately equal
up to a scale factor, such that DOG(x)=LOG(x)  constant.
Specifically, by considering the values of the two functions
at the origin x = 0, it is easy to see that

kxk2

e 22 (1 6) (14)
:

One reason to prefer the DOG approximation over LOG

is that there exist fast techniques for generating Gaussian
convolution pyramids. We prefer the DOG operator in the
present context because it is more clearly related to the
mean-shift algorithm using a Gaussian kernel, a connection
that will be explored in the next section.
It is interesting to note that the magnitude of the DOG
operator is already invariant across scales, thus the operator
does not have to be normalized by scale parameter 2 , as the
LOG filter does. That is, if the convolution of image f (x)

We adapt feature scale selection theory to create a mechanism for adapting mean-shift kernel size while tracking
blobs through changes in scale. Intuitively, we will track a
blobs location and scale by using the mean-shift algorithm
to track the local maximum (Eq 9) that represents the blob
feature in the scale space generated by the DOG operator.
Define a 3D shadow kernel H (x; ) with two spatial dimensions x and one scale dimension . At any given scale
0 , the 2D marginal kernel H (x; = 0 ) will be a spatial
filter DOG(x; 0 ). The set of DOG filters at multiple scales
form a scale-space filter bank, which is convolved with a
sample weight image where each pixel is proportional to
the likelihood that it belongs to the object being tracked.
See Figure 3. The results of the DOG filters at multiple
scales are then convolved with an Epanichikov shadow kernel in the scale dimension. The result is a 3D scale space
representation in which modes represent blobs at different
spatial positions and scales. We want to design a mean-shift
filter to track modes through this scale space representation.
More formally, define a set of kernel scales around the
current scale 0 as

fs = 0  bs

for

n  s  ng

(15)

where b > 1 specifies the base of a logarithmic coordinate

scale and n limits the range of scales to explore around the
current scale 0 . We typically choose b = 1:1 and n = 2.
The shadow kernel in the scale dimension is the Epanichnikov kernel
(16)
Hs (s) = 1 (s=n)2

Figure 3: Shadow kernel for mode-seeking in scale space.

This 3D kernel takes the form of a filter bank of centered
DOG spatial kernels at different scales, convolved with a
1D Epanichnikov kernel across the scale dimension.
A filter bank of spatial shadow kernels is also defined, one
for each scale s . Each is defined as a DOG kernel around
the current object location x0
Hx (x; s) = DOG(diag(1=w; 1=h)(x

x0); s )

(17)

The scale factors w and h are introduced to allow elliptical

spatial kernels (the DOG operator is circular). Finally, the
3D scale space is generated from a sample weight image
w(x) as
(18)
E (x; s) = Hs (s) Hx (x; s)w(x)
s

We now wish to design a mean-shift procedure that finds

local modes in the scale space defined by the 3D convolution of Eq. 18. We simplify by considering two interleaved
mean-shift procedures, one spatial and one in scale, that iterate to find scale space modes. To find the mean-shift kernels Ks and Kx in image space that correspond to shadow
kernels Hs and Hx in scale space, we use the kernel-shadow
relationship constraint of Eq. 3.

4.1 Spatial kernel

First, we treat the current scale 0 as constant, and consider
the spatial shadow kernel Hx (x; s). Since the DOG operator
is a linear combination of Gaussians, and since the Gaussian
kernel is its own shadow, we should expect the mean-shift
kernel Kx (x; s) to also be a linear combination of Gaussians.
Indeed, after applying Eq. 3 with respect to r = (x x0 )2 we
find that
Kx (x; s) = G(x; x0 ; 1 )=21

G(x; x0 ; 2 )=22

(19)

with 1 = s =1:6 and 2 = 1:6s . This kernel is also a

DOG kernel, but not one that approximates a Laplacian it
has a more pronounced peak and a shallower negative basin
around the peak. The mean-shift kernel corresponding to
E (x; s) in Eq 18, for fixed 0 is therefore
KE (x) = Hs (s)Kx (x; s)
s

(20)

This is a linear combination of filter bank responses, with

the highest weight at the current scale 0 , and weights dropping off as the scale of the filter increases or decreases.
One snag is that each Kx (x; s) is a non-convex kernel that
contains negative weights, and it therefore violates the assumptions that are typically placed on mean-shift kernels.
The practical difficulty of this is that during computation
of the mean-shift offset (Eq 1), some pixel values will be
multiplied by negative kernel weights. However, we have
already dealt with this problem in Section 2, where we considered the related problem of negative sample weights. The
solution is the same: during computation of the mean-shift
offset vector, we divide by the sum of the absolute values of
all weight terms, instead of just the sum of the terms. With
this modification, which was shown in Eq. 7, mean-shift
with the Kx (x; s) kernel is well-behaved.

4.2 Scale kernel

Now consider holding the location x0 fixed, and looking
for the best scale s . The mean-shift kernel corresponding to the Epanichnikov shadow in the scale dimension
will be a flat kernel. This kernel is applied to the values
x Hx (x; s)w(x), for the range of scales n  s  n. Intuitively what is happening is that, for a given spatial location x0 , a filter bank of DOG operators is applied at varying
scales, centered on that location, and then mean-shift is performed on the 1D array of results to locate the scale mode.

4.3 Algorithm summary

To summarize, we use two interleaved mean-shift procedures to track modes in scale space, which represent the
spatial location and scale of blobs in an image. The algorithm proceeds as follows
1. Input is a sample weight image w(a) and an estimate
of the blobs current scale space location, represented by
(x0 ; 0 ), a parameter vector that combines both spatial and
scale terms.
2. Hold 0 fixed, and perform a spatial mean-shift procedure using the equation
x =

s Hs (s) a Kx (a x0; s)w(a)(a x0 )

s Hs (s) a jKx (a x0; s)w(a)j

(21)

where n  s  n defines a range of scales s as in Eq 15.

3. Let x0 now represent the value that the spatial mean-shift
procedure in step 2 converges to. Holding x0 fixed, perform
a mean-shift procedure using the equation
s0 =

s x Hx (x; s)w(x)s
:
s x Hx (x; s)w(x)

(22)

This procedure is iterated until convergence. Set the new

0
value of 0 to 0  bs .

4. Iterate by interleaving steps 2 and 3 until both ks0 k < s

and kx k < x .

5. Examples
Figure 4 will help to better understand the motivation behind the scale selection approach presented in this paper.
The top of Figure 4 shows three blobs (squares) of different
sizes. The bottom of Figure 4 shows one slice through the
3D scale space generated by convolution with the DOGEpanichnikov filter bank defined in the previous section.
The modes in this scale space clearly localize each of the
blobs both spatially, and in the scale dimension, as expected
from the feature scale selection theory of Lindeberg. In the
previous section we designed an interleaved mean-shift procedure that can find/track these modes without having to
explicitly generate the scale space.

used in [4], who show that these weights are related to maximizing the similarity of histograms di and mi as measured
by the Bhattacharyya coefficient.
Sample frames of the tracking results from three different algorithms are presented. Figure 5A shows tracking results from the classic mean-shift algorithm without
any changes of scale. The person is successfully tracked
throughout the sequence, but the fixed scale of the tracking kernel provides poor localization of the centroid of the
person as their size increases near the end of the sequence.
Figure 5B shows tracking results when the scale is
adapted using the approach suggested in [4]. At each iteration, the mean-shift algorithm is run three times, once with
the current scale, and once with window sizes of plus or minus 10 percent of the current size. For each, the color distribution observed within the mean-shift window after convergence is compared to the model color distribution using the
Bhattacharyya coefficient, and the window size yielding the
most similar distribution is chosen as the new current scale.
We see that the window quickly shrinks too much, a common failure of this scale selection approach. At roughly a
third of the way through the sequence, the tracking is lost.
Finally, Figure 5C shows the results of our kernel size
selection method, described in the last section. The person is consistently tracked, both in image location AND in
scale. Note in particular the correct adaptation of kernel
scale as the persons size rapidly expands near the end of
the sequence.

6. Conclusion

Figure 4: Intuitive illustration of scale space modes. Top:

three blobs (squares) of different sizes. Bottom: scale space
generated by the DOG-Epanichnikov filter bank defined in
this paper. Each blob is well localized both spatially and in
the scale dimension by a mode in this scale space.
Figure 5 shows blob tracking results on a real image sequence of a person walking around a test trajectory. The
sequence comes from a database of gait test sequences collected at NIST as part of the DARPA Human Identification
program. An image rectangle is selected by hand in the first
frame, and a color histogram mi is acquired to model the
color distribution within this rectangle. Sample weight images are generated with respect to a data histogram di of
colors within the current tracking window, with each pixel
of color i being set to mi =di . This is the same method

We have addressed the problem of choosing the correct

scale at which to track a blob using the mean-shift algorithm. We start with Lindebergs theory of blob scale detection using differential scale-space filters, and define an
efficient mean-shift procedure that tracks the modes in this
scale space. Each mode represents the spatial location and
scale of an image blob.
Conceptually, the scale space is generated by convolving a filter bank of spatial DOG filters with a sample weight
image. The results are then convolved with an Epanichikov
kernel in the scale dimension. Explicit generation of these
convolutions in order to find the modes would be very expensive. However, by exploiting the relationship between
shadow kernels and mean-shift kernels, we can derive designer mean-shift kernels that find the modes much more
efficiently. A two-stage mean-shift procedure is developed
that interleaves spatial and scale mode-seeking to track a
local mode in scale space. The result is stable and natural
selection of an appropriate kernel size for appearance-based
image blob tracking.

(A)

(B)

(C)
Figure 5: Tracking examples. (A) Using a fixed-scale mean-shift kernel. The person is tracked through the sequence, but
localization is poor when the persons size increases. (B) Using the plus or minus 10 percent scale adaptation method (see
text). The kernel soon shrinks too much, leading to tracking failure. (C) Using the scale-space mode-tracking method
presented in this paper. The person is tracked well, both spatially and in scale.

References
[1] Bradski, G.R., Computer Vision Face Tracking for
Use in a Perceptual User Interface, IEEE Workshop
on Applications of Computer Vision, Princeton, NJ,
1998, pp.214-219.
[2] Bretzner, L. and Lindeberg, T., Qualitative Multiscale Feature Hierarchies for Object Tracking, Journal of Visual Communication and Image Representation, Vol 11(2), June 2000, pp.115-129.
[3] Cheng, Y., Mean Shift, Mode Seeking, and Clustering, IEEE Trans. Pattern Analysis and Machine Intelligence, Vol 17(8), August 1995, pp.790-799.
[4] Comaniciu, D., Ramesh, V. and Meer, P., Real-Time
Tracking of Non-Rigid Objects using Mean Shift,
IEEE Computer Vision and Pattern Recognition, Vol
II, 2000, pp.142-149.
[5] Comaniciu, D., Ramesh, V., Meer, P., The Variable
Bandwidth Mean Shift and Data-Driven Scale Selection, International Conference on Computer Vision,
Vol I, pp.438-445.
[6] Fukanaga, K. and Hostetler, L.D., The Estimation
of the Gradient of a Density Function, with Applica-

tions in Pattern Recognition, IEEE Trans. Information Theory, Vol 21, 1975, pp.32-40.
[7] Hildreth, E.C., The Detection of Intensity Changes
by Computer and Biological Vision Systems, Computer Vision, Graphics and Image Processing, Vol
22(1), April 1983, pp.1-27.
[8] Lindeberg, T., Feature Detection with Automatic
Scale Selection, International Journal of Computer
Vision, Vol 30(2), November 1998, pp.79-116.