REGIONS OF INTEREST FOR ACCURATE OBJECT DETECTION

P. Kapsalas, K. Rapantzikos, A. Sofou, Y. Avrithis

Image Video and Multimedia Systems Laboratory,
Department of Electrical & Computer Engineering,
Athens 15780, Greece
{pkaps, rap, natasa, iavr}@image.ntua.gr
Abstract
In this paper we propose an object detection approach that
extracts a limited number of candidate local regions to guide
the detection process. The basic idea of the approach is that
object location can be determined by clustering points of
interest and hierarchically forming candidate regions
according to similarity and spatial proximity predicates.
Statistical validation shows that the method is robust across
a substantial range of content diversity while its response
seems to be comparable to other state of the art object
detectors.
1.

INTRODUCTION

The rapidly expanding research and analysis on object

detection and recognition imposes the use of reliable
machine vision systems. Recent object detection approaches
fuse detection results obtained by robust part detectors, thus
eliminating noise artifacts and resolving occlusion
phenomena. However, object detection remains a
challenging problem due to the wide range of variability in
the monitoring conditions as well as to the diversities in
scale, location, orientation (up-right, rotated), and pose that
real objects adopt.
The goal of an efficient object detector is the accurate
determination of objects location, extent and shape. To
efficiently approximate these parameters, there are many
challenges that should be faced. These are mainly associated
with the variability of poses and orientations that the object
of interest can adopt. Monitoring parameters such as
illumination conditions, camera movements, aliasing and
camera specifications also correspond to issues that should
be considered
Through this section we provide a brief review of
important existing object detection techniques in static
images and video. However, a full review of the object
detection literature is beyond the scope of this paper. In an
effort to distinguish the existing techniques, we can classify
them into four broad categories, with respect to the
information that they consider. However, some overlapping
between categories can occur.

Bottom-Up Feature-Based Approaches aim to find

structural features that exist even when the pose, viewpoint,
or lighting conditions vary, and then use them in the
detection procedure. Davis & Sharma [1] proposed a
contour-based method to detect humans in widely varying
thermal images. At first, regions of interest (ROIs) are
determined through statistical background subtraction.
Subsequently, gradient information within each region is
extracted and used to form a contour saliency map.
Morphological operations are also employed to fill broken
boundaries. The combination of gradient and contour-based
features is an additional class of descriptors that has been
extensively used. Their efficacy relies on the robustness of
gradient-based descriptors to illumination changes and noise
induction. Such features are frequently selected to represent
objects boundaries.
The Scale Invariant Feature Transform (SIFT) and
shape context have been extensively used for person/face
localization [2], [3], [4], [5]. SIFT was initially proposed by
Lowe [2], [3] and works by combining a scale invariant
region detector and a descriptor based on the gradient
distribution in the detected regions. Geometric histogram [4]
and shape context [5] implement the same idea and are very
similar to SIFT. In particular, shape context is identical to
SIFT descriptors but is based on edges extracted by the
Canny [6] detector where the location is quantized into nine
bins of a log-polar coordinate system.
Another approach also based on gradient orientation
distributions is proposed by Dalal & Triggs [7]. A dense
grid of Histograms of Oriented Gradients (HoG) is
considered and computed over blocks of specific extent.
This representation has proved to be powerful enough to
classify humans using a linear SVM. However, it can only
process images of limited dimensions and at specific frame
rates using a very sparse scanning methodology that
evaluates roughly 800 detection windows per image. Zhu et
al. [8] speed-up Dalals approach by increasing the number
of detection windows. Thus, they combine a cascade of
rejectors approach with the HoG features. To overcome fast
rejection problems, they induce the use of fixed size blocks.

They also use a much larger set of blocks varying in sizes,

locations and aspect ratios. The best blocks suited for
detection are selected via an AdaBoost relying on a rejectorbased cascade. Visual attention has also been used in several
object detection approaches in order to account for motion
information and conscious search in images. Rapantzikos
and Tsapatsoulis [9] have built a robust method of
enhancing the accuracy of face detection schemes through a
visual attention architecture.
Top-Down Knowledge-Based Methods are rule-based
approaches that encode knowledge of what constitutes an
object of interest. Knowledge-based methods have the
fundamental advantage of attaining to reduce false positive
instances of detection by eliminating them through the
verification step. Motion characteristics provide important
information for both determining regions of interest and
assessing whether an objects motion features resemble to
human motion [10], [11],[12]. Viola and Jones [11] have
proposed a state of the art human detection approach, which
considers prior knowledge on the persons motion and
appearance. No separate mechanisms of tracking,
segmentation and alignment are supported. The system
works by simply selecting the feature set, the scale of the
training data and the scales used for detection. The training
process uses AdaBoost to select a subset of features and
construct the classifier. The classifier consists of a linear
combination of the selected features. Viola and Jones [13]
have also proposed a cascade of classifiers architecture to
reduce the computational cost.
Template Matching Methods use standard patterns of
objects/object parts to describe the object globally or as
distinct parts. Correlations between the input image and
patterns subsequently computed for detection. Gavrila [12]
propose a human detection scheme that segments
foreground regions and extracts the boundary. Then the
algorithm searches for humans in the image by matching
edge features to a database of templates of human
silhouettes. The matching is realized by computing the
average Chamfer distance between the template and the
edge map of the target image area. Wren et al. [14] describe
a top-down person detector based on template-matching.
However, this approach requires domain specific scene
analysis. Castillo and Chang use fast template matching as a
focus of attention [14] and the algorithm proceeds by
discarding locations where there is no silhouette that
matches the human body.
In the appearance-based methods, the models (or
templates) are learned from a set of training images. These
learned models are then used for detection. Appearancebased methods [16] rely on techniques from statistical
analysis and machine learning to find the relevant
characteristics of images containing objects. The learned
characteristics are in the form of distribution models or
discriminant functions that are consequently used for
detection.

In contrast to all techniques described above,

Integration of Parts detectors fuses the detection results
derived by robust part-based detectors. Forsyth and Fleck
[17] introduced body plans for finding people in general
configurations. Ioffe and Forsyth [18] then assembled body
parts with projected classifiers or sampling. However, the
aforementioned methods rely on simplistic body part
detectors. The employed representation models body parts
as bar shaped segments. An improvement on the modeling
of body part relations is given by Mikolajczyk et al. [19],
where local orientation position features are extracted by
gradient and Laplacian based filters. The spatial layout of
the features, together with their probabilistic co-occurrence
captures the appearance of the parts and their
distinctiveness. The features with the highest co-occurrence
probabilities are learnt using AdaBoost and the detected
parts are combined with a joint probabilistic body model.
The features deployed in [19] have proven to describe the
shape better than prior descriptors. Parts-based object
detection and recognition systems were further extended in
[20] and [21], where several methodologies have been
developed based on either bottom-up or top-down
considerations. Bottom-up approaches were used to group
together body parts found throughout a sequence while in
top-down approaches, human models are built automatically
from convenient poses. The system in [20] has been also
used to track humans by detecting the learned models in
each frame.
The majority of object detection methodologies
approach the localization issue through exhaustive search
over the image. Thus, false positive artifacts are eliminated
at the cost of distorting the object extent and shape. In this
paper we try to tackle with the localization problem through
considering a set of interest points describing regions that
represent transition areas with high probability. These points
are subsequently considered to guide the overall detection
process. The basic idea is that interest points are located on
objects boundaries or on areas of abrupt changes on the
background structure.
In practice we determine interest points locations
through Harris corner detector [22] thus enhancing points
characterized by high gradient values along all directions.
Points grouping (clustering) takes place by measuring
distances on low level visual descriptors. The derived
clusters are considered as local image structures
representing the content of the local neighborhood.
Structuring the paper, section 2 presents a brief
overview of the systems functionality while sub-sections
2.1 through 2.4 analyze the functionality provided by each
distinct module. The experimental setup and the results are
presented and evaluated in sub-sections 3.1 and 3.2
respectively. Systems performance issues are also
considered in terms of both statistical and visual analysis.
Finally, section 4 draws the conclusions and discusses drifts
of further research.

2.

PROPOSED FRAMEWORK

The proposed methodology approaches the problem of

object detection by initially extracting distinctive invariant
features from images that are in turn used for region
grouping. The features are extracted from local keypoints
derived by the Harris corner detector [7] and are associated
with point locations that have large gradients in all
directions at a predetermined scale. The selected features
reflect color and texture information evaluated on local
regions and have proven to be invariant across a substantial
range of image deformations and illustration conditions.
Apart from considering the distributions of color and
intensity metrics the employed features also account for
spatial dependencies of intensity levels.
After the feature points have been determined they are
grouped via an unsupervised clustering approach. An
efficient clustering approach should be characterized by
accurate discovery of clusters of arbitrary shape (shape of
clusters in spatial databases may be spherical, drawn-out,
linear, elongated etc). Moreover, feature points clustering
should provide good efficiency when applied on large
databases (i.e. on databases of significantly more than just a
few thousand objects) while it should also be invariant to
prior knowledge on the application data.
In our approach localization of objects initially involves
formation of neighborhoods of clusters according to spatial
proximity distance. Subsequently, the clusters belonging to
each neighborhood are considered in combinations to check
the objects occurrence at each specific location. The final
step involves the use of a cascade of boosted classifiers.
The overall functionality supported in our method is
illustrated in Fig. 1.
Detect Points Of
Interest

Feature Extraction
Unsupervised
Clustering (DBScan)
Check Combinations of
ROIs at all Spatial
Configurations

Verification
Stage

Figure 1: Object detection system in diagram form

2.1. Interest Point Extraction and Local Features

Our detection methodology initially considers the point
locations extracted by the Harris corner detector [22]. These
are determined by evaluating the autocorrelation function of
the 2D visual signal (image) within a spatial region of
predetermined extent. The autocorrelation function measures
the local changes of the patches shifted by a small amount in
different directions. Given a shift (x, y) and a point (x, y),
the autocorrelation function is defined as,
c ( x, y ) = I ( xi , yi ) I ( xi + x, yi + y )
W

(1),

where I denotes the image function and (xi, yi) are the points
in the window W (Gaussian) centred on (x, y). The shifted
image is approximated by a Taylor expansion truncated to
the first order terms.
x
I ( xi + x, yi + y ) I ( xi , yi ) + I x ( xi , yi ) I y ( xi , yi ) (2)
y
Substituting (1) into (2) we get,
x
c ( x, y ) = [ x y ] C ( x, y )
(3),
y
where matrix C(x,y) captures the intensity structure of the
local neighborhood.
Let 1, 2 be the eigen-values of matrix C(x,y). The
eigen-values form a rotationally invariant description. The
presence of a key-point depends on the relation between 1
and 2. Low levels of 1, 2 are associated with flat
autocorrelation function that corresponds to an image area
with no abrupt change in any direction. If only one of the
eigen-values is high and the other is low, that implies a ridge
shaped auto-correlation function. Such ridges in the autocorrelation function are associated to the occurrence of
transition boundaries between different surfaces. High levels
of 1, 2 are reflected to sharp peaks on the autocorrelation
function and then shifts in any direction will result in a
significant increase that indicates a corner.
Point detection is followed by spatial filtering to
eliminate the density of feature points within local
neighborhoods. Equation (4) provides a mathematical
description of the points filtering operation. Set S
corresponds to the set of locations obtained after an initial
selection of interest points.
S = { p H : p q < , q H}
(4),
where p denotes the point under consideration, H the set of
points obtained by the Harris corner detector and a
parameter denoting the radius of the area being checked.
The size of is chosen to be significantly smaller than the
patch considered for extracting visual descriptors. The
underlying idea of selecting the value of is to suppress the
number of points co-occurring in the same spatial
neighborhood while also enabling accurate representation of
local image structures.
Keypoints localization is followed by a features
extraction procedure. At this step, we consider local square
patches of limited extent (significantly larger than ) and
evaluate colour and texture descriptors to represent local
areas. Thus, the three most significant components of
dominant colour are selected to reflect the local colour
content while a statistical descriptor of texture was used to
estimate texture variations within images sub-regions. More
specifically, at each keypoint location the gray-level cooccurrence matrix is extracted and some features expressing
the contrast and homogeneity, within these areas, are
evaluated. The co-occurrence matrix displacement vector is
selected so as to represent the spatial arrangement of

intensity levels while suppressing the induction of intensity

peaks associated to noise occurrence.
The texture features considered through this work are
the Harralick descriptors of texture and they actually
measure the randomness of gray levels distribution. The
mathematical expressions of these features are provided
through the subsequent equations.
Energy : F1 = P 2 ( i, j )
i

After the clusters have been formed, they are examined

according to spatial proximity criteria to form spatial
neighborhoods in the image plane. The clusters within each
neighborhood are subsequently checked in combinations to
derive candidate regions. The latter are selected to provide
local area description and are considered for feature
representation. Figure 2 summarizes the operations deployed
throughout our methodology:

Entropy : F2 = P ( i, j ) log ( P ( i, j ) )
i

Homogeneity : F3 =
i

P ( i, j )
1+ i j

and

Contrast : F4 = ( i j ) P ( i, j ) ,
2

Where P(i,j) denotes the probability of finding pixels with

intensity values i, j at spatial arrangements similar to the one
defined by the displacement vector.
Feature normalization is also an issue of great concern
as it influences the classifiers performance. In this work,
L2square-normalization is considered for normalizing the
input vectors. The same distance metric is also used by the
clustering algorithm [23] for discriminating clusters in the
feature space.

(a)

(b)

2.2. Interest Point Clustering

The clustering approach groups feature points according to
similarity and spatial proximity criteria. In practice, the
employed algorithm called Density Based Algorithm for
Discovering Clusters in Spatial Databases with Noise
(DBScan) [23] has proven to be powerful enough to
discover clusters of arbitrary shape while requiring minimal
domain knowledge. The algorithm considers density of
feature points as the key aspect for forming meaningful
classes in the feature space. Adaptation of the clustering
algorithm requires modifying the maximum neighbourhood
radius (Eps) and the minimum number of points (MinPts) in
an Eps-neighborhood of that point.
The functionality of the DBScan algorithm can be
summarized in the following steps.

(c)
(d)
Figure 2: (a) Points of interest (marked in red), (b) Clusters derived after
applying the DBScan (yellow rectangles enclosing interest points), (c)
Neighborhoods of clusters on the image, (represented as blue ellipses), (d)
candidate regions (enclosed in the light-green rectangles)

As it can be observed the localization process initially

involves the determination of neighbourhoods of clusters
(represented by blue-coloured ellipses in Fig. 2 (c)) and at
the next step, the clusters contained within each
neighbourhood are considered in combinations to obtain the
candidate regions (light green rectangles in figure 2(d)).

1. Select arbitrarily a p point in the feature space.

2.3. Feature Extraction

2. Select all points that are density reachable from p

w.r.t Eps and MinPts.

The feature extraction approach is based on evaluating wellnormalized local features of image gradient orientations in a
dense grid [7]. The basic idea is that local object appearance
and shape can be characterized rather well by the
distribution of local intensity gradients. The features are
selected to be robust to illumination changes and imaging
conditions and this is satisfied by encoding the objects
boundary orientation and discarding information relative to
the local colour or intensity. The histogram representation
also enhances the methods robustness to rotation changes.
In practice, the implementation involves dividing the
image window into small spatial regions (cells) and for
each cell accumulating a local 1-D histogram of gradient

3. If p is a core point, a cluster is formed.

4. If p is a border point, no points are density reachable
from p and DBScan visits the next point of the
database.
5. Continue the process until all points have been
processed.
Where the terms density reachable and directly-density
reachable are associated with the radius of the points Epsneighborhood and the minimum number of points (MinPts)
in this neighborhood [23].

directions or edge orientations over the pixels of the cell.

More specifically, each pixel calculates a weighted vote for
an edge orientation histogram channel based on the
orientation of the gradient element centred on it. The votes
are evaluated into orientation bins over cells. The combined
histogram entries form the representation of possible object
boundaries. The vote is a function of the gradient magnitude
2
at the pixel (e.g. I ( x, y ) , I ( x, y ) or I ( x, y ) ). For
better invariance to illumination, shadowing, etc., it is also
useful to apply contrast normalization of the local features
before using them [7]. This can be done by accumulating a
measure of local histogram energy over larger spatial
regions (blocks) and using these results to normalize all
cells within the block. The histogram of oriented gradients
technique has the advantage of capturing characteristic edge
or gradient structure information, which can represent local
shape with an easily controllable degree of invariance to
local geometric and photometric transformations. Thus,
translations or rotations make little difference if they are
much smaller than the local spatial or orientation bin size.
An overview of the feature extraction approach is illustrated
in Fig. 3.
Intensity
Normalization

Compute Gradients

Weighted Vote into

Spatial & Orientation
Cells

Verification Stage
(AdaBoost)

Collect HOGs over

Spatial Regions

Normalize Contrast
Over Spatial Regions

2.4. Classification
For the candidate regions defined in 2.2, we evaluate the
feature set representing the local image content and employ
a cascade of boosted classifiers [13] at the verification stage.
The underlying idea in this approach is that smaller and
therefore more efficient boosted classifiers can be
constructed in order to reject many of the negative areas
while detecting almost all positive instances. Simpler
classifiers are used to reject the majority of regions of
interest before more complex classifiers are called upon to
achieve low false positive rates. The overall form of the
detection process is that of a degenerate decision tree, also
called a cascade (see Fig. 4).
T

1
F
Reject

...

N-1

...

EXPERIMENTAL RESULTS

3.1 Experimental Setup

Our detection framework is evaluated upon a broad range of
images depicting the objects of interest in divert poses and
at different illustration conditions so as to resemble the
objects occurrence in real scenes. More specifically, the
person detection framework considers a part of the
PASCAL dataset as training set. The training set contains
2380 positive instances and over 3500 negative examples.
Based on this content we have trained three individual
classifiers focusing on the full, upper and lower body
respectively. However, in the experimental results presented
in the following section only the full body classifier is
considered so as the results to be comparable with other
person detection approaches. The testing set used for
evaluating the systems performance contains over 800
images depicting frontal or backward views of human
bodies. Regarding the cars and airplanes content, we have
produced training sets from the images provided by the
PASCAL databases. More specifically, the car detector
training set consists of approximately 1800 positive
examples and around 2200 negatives while the test set
involves 2200 images depicting both negative and positive
examples. The corresponding training set for the airplanes
contains 1100 positive images and 1300 negative examples.
3.2 Results

Figure 3: Overview of the feature extraction approach.

Candidate
Regions

3.

Further
Processing

F
Reject

Figure 4: Schematic depiction of the detection cascade. A series of

classifiers are applied to regions of interest. The initial classifier eliminates
a large number of negative examples with very little processing.
Subsequent layers eliminate additional negatives but require additional
computation.

The testing framework initially validates the potential of our

approach towards the effective determination of the
topology and extent of objects of interest. A second
objective is to statistically estimate the systems efficiency
in detecting objects of interest across a great diversity of
monitoring conditions. Figs. 5(a)-(l) depict example of
objects detected through our system. Visual inspection of
the results indicates that the methodology is quite efficient
in determining the objects of interest at their exact location
and extent while eliminating the induction of false positives.
More specifically, Fig. 5(a)-(d) illustrate the performance in
detecting persons at variable scales and under different
monitoring conditions. The visual results denote that our
methods performance is comparable to other state of the art
object detectors. Moreover, its response is closely related to
the discriminability provided by the feature set that is used
to represent the objects. The robustness against scale
changes and its efficacy at suppressing noise artifacts is
strongly associated with the image representation.
For the statistical evaluation a number of parameters are
being varied and in each case the methods response is
evaluated in terms of its potential to detect objects of
interest at their actual extent and location.
Through the evaluation procedure, pixels that co-occur
in both the detected and the ground truth image are
considered as true positive (TP) instances of detection.

Similarly, true negative (TN) instances correspond to image

locations classified as background by both the detector and
the ground truth, while false positive (FP) and false
negatives (FN) are determined accordingly. In this paper, we
consider the Receiver Operating Characteristic (ROC)
curves as robust measures for evaluating the algorithmic
performance. The ROC curves provide information on the
tradeoff between the algorithms specificity (SP) and
sensitivity (SE) [24].
TN
TN
(6)
SP =
=

(TN + FP )

Similarly the sensitivity value is defined as

SE =

TP

(TP + FN )

TP
P

(5)

The statistical evaluation procedure involves 5-fold

cross-validation to guarantee more robust estimation of
response. Fig. 6, 7, 8 and 9 provide the statistical curves
representing our approach performance. Through the ROC
curves, we estimate the trade-off between the systems
sensitivity and specificity when adjusting two important
parameters of the overall detection methodology. In
particular, Fig. 6 illustrates the systems response when the
varying parameter is the cell size while Figs. 7 and 8
illustrate performance variations when the DBScan
parameters are modified.
A brief study of the derived results illustrated in Fig. 5
indicates that our object detection technique seems to be
more robust when detecting cars while the performance of
person and airplane detection follows. Particularly, airplane
detection seems to have significantly lower performance.
This is closely related to the extent of the available training
set as well as to the degree to which the training set
represents real scenes of airplanes occurrence. The diversity

of the testing set is also a significant factor affecting the

overall performance.
The effect of DBScan density parameter (Eps) and
minimum points (MinPts) on detection performance is
illustrated in figures 7 and 8 respectively. As it is observed,
the detector seems to provide better results when applied to
cars and to airplanes detection problem than humans. This
can be explained by considering that people tend to be
encountered at more complex scenes. This introduces
limitations to the performance of the clustering module,
which in turn affects the localization process.
Further comparisons between Fig. 6, 7 and 8 indicate
that the variation of the DBScan parameters leads to more
abrupt changes on the performance curves than the cell size.
This can be explained by considering that the region
clustering guides the overall detection procedure. Further
comparison between Figs. 7 and 8 indicates that
performance is more sensitive to the MinPts parameter than
Eps. This can be explained by considering that the minimum
number of points affects the overall clustering procedure as
it may lead to very large (very low) number of clusters,
which in turn obstructs the determination of candidate
regions. Experimental and statistical evaluation led to the
assessment that the optimum values for the parameters being
modified are: Eps=0.6, MinPts=4 and cell size=8x8. These
values were also considered in the extraction of the ROC
curves.
Finally, Fig. 9 compares the performance of our
approach to the approach presented in [7]. The comparison
is based on the potential of the two methods to detect
persons on the same dataset (INRIA). It can be observed,
that our approach attains a slightly better response at the
critical area while both approaches present almost the same
performance for extreme values of specificity.

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)
(j)
(k)
(l)
Figure 5: Results of our object detection system (a)-(d): Person detection under different scales and poses, (e)-(h) Detection of cars, (i) (l): Detection of
airplanes.

It can also be seen that our detector is invariant to the

image size as the overall procedure is guided by the
determination of regions of interest.
Receiver Operating Characteristic Curves (ROC)
0.9
0.8
0.7

Sensitivity

0.6
0.5
0.4
0.3
Cars
AirPlanes
Persons

0.2
0.1

0.2

0.4
0.6
1Specificity

0.8

Figure 6: ROC curves obtained by modifying the cells size.

Receiver Operating Characteristic Curves (ROC)
0.9
0.8
0.7

Sensitivity

0.6
0.5
0.4
0.3
Cars
AirPlanes
Persons

0.2
0.1

0.2

0.4
0.6
1Specificity

0.8

Figure 7: ROC curves derived by modifying the DBScan Eps parameter.

Receiver Operating Characteristic Curves (ROC)
0.9
0.8
0.7

Sensitivity

0.6
0.5

CONCLUSIONS AND FUTURE WORK

Throughout this paper we proposed an object detection

approach and evaluated it in real scenes. The approach
initially considers points of interest to determine locations
close to boundaries. Such points are then grouped according
to similarity and spatial proximity criteria. Feature points
grouping takes place through an unsupervised clustering
approach across a large range of domains (generic detector).
The overall localization problem is achieved by forming
neighborhoods of clusters. The clusters belonging to each
neighborhood are checked in combinations to form
candidate regions and in turn to assess the exact location of
objects occurrence.
The response is estimated in terms of both visual and
statistical evaluation. The test sets are produced so as to
depict objects at variable scales and poses. Visual inspection
illustrated that the proposed methodology provides accurate
determination of the object location and extent. Statistical
evaluation also verified that the approach has a performance
that is comparable to state of the art object detectors. The
overall approach has no limitations related to the image size
and the object position.
Although our detection methodology provides quite
accurate results, there is still room for further optimization.
One of our initial objectives is to evaluate the methods
efficiency in terms of detecting objects of interest on other
datasets. The TRECVID content is a possible data set for
comparisons to other detectors although it does not contain
local information in its ground truth. A future direction is
the investigation of alternative features (e. g Haar features).
The overall detection procedure could also be strengthened
by using contour-based information to guide grouping of
clusters.

0.4
0.3
0.2
Cars
AirPlanes
Persons

0.1
0

0.1

0.2

0.3

0.4
0.5
1Specificity

0.6

0.7

0.8

0.9

Figure 8: ROC curves derived by modifying the DBScan MinPts

parameter.

0.8
0.7
0.6

5.

REFERENCES

[2] D. G. Lowe, Distinctive Image Features from Scale-Invariant

Keypoints, Intl J. Computer Vision, vol. 2, no. 60, pp. 91-110,
2004.

0.5
0.4
0.3
0.2
0.1

ACKNOWLEDGEMENTS
This research work was supported in part by the European
Commission
under
contracts:
FP6-2005-IST-5
IMAGINATION, FP6-027685 MESH and FP6-26978 XMedia.

[1] J.W. Davis, V. Sharma, Robust Background Subtraction for

Person Detection in Thermal Imagery, In: Proc. of the IEEE
Conference on Computer Vision and Pattern Recognition
Workshops (CVPRW04), vol. 8, pp. 128-132, 2004.

Receiver Operating Characteristic Curves (ROC)

0.9

Sensitivity

4.

Our Detector
HOGsTriggs
0

0.1

0.2

0.3

0.4
0.5
1Specificity

0.6

0.7

0.8

0.9

Figure 9: Comparative study of the performance curves representing our

method with the approach described in [7].

[3] D.G. Lowe, Object Recognition from Local Scale-Invariant

Features, Proc. Seventh Intl Conf. Computer Vision, pp. 11501157, 1999.
[4] Ashbrook, N. Thacker, P. Rockett, C. Brown, Robust
Recognition of Scaled Shapes Using Pairwise Geometric

Histograms, In Proc. Of the 6th British Machine Vision

Conference, pp. 503-512, 1995.
[5] S. Belongie, J. Malik, J. Puzicha, Shape Matching and
Object Recognition Using Shape Contexts, IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 2, no. 4, pp. 509522, 2002.
[6] J. Canny, A Computational Approach to Edge Detection,
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 8, no.
6, pp. 679-698, 1986.
[7] N. Dalal, B. Triggs, Histograms of Oriented Gradients for
Human Detection, In: Proc. of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition (CVPR
2005), vol. 1, pp. 886-893, 2005.
[8] Q. Zhu, S. Avidan, M.C. Yeh, K.T. Cheng, Fast Human
Detection Using a Cascade of Histograms of Oriented Gradients,
In: Proc. of the 2006 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR06), vol. 1, pp.
1491-1498, 2006.
[9] K. Rapantzikos, N. Tsapatsoulis, Enhancing the robustness
of skin-based face detection schemes through a visual attention
architecture, In: Proc of the IEEE International Conference on
Image Processing, Genova, Italy, 2005.
[10] J. K. Aggarwal and Q. Cai, Human motion analysis: A
review, Computer Vision and Image Understanding, vol. 73, no.
3, pp. 428440, 1999.
[11] P. Viola, M. J. Jones, and D. Snow, Detecting pedestrians
using patterns of motion and appearance, In Proc. of the 9th
International Conference on Computer Vision, vol. 1, pp. 734-741,
2003.
[12] D.M. Gavrila and V. Philomin, Real-time object detection
for smart vehicles, IEEE Computer Society Conference on
Computer Vision and Pattern Recognition Workshops (CVPR), pp.
8793, 1999.
[13] P. Viola and M. Jones, Rapid object detection using a
boosted cascade of simple features, In Proc. of the IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
vol. 1, pp. 511-518, 2001.
[14] C. R. Wren, A. Azarbayejani, T. Darrell, and A. Pentland,
PFinder: Real-time tracking of the human body, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol.
19, no. 7, pp. 780-785, 1997.
[15] C. Castillo and C. Chang. An approach to vision-based
Iberian
person detection in robotic applications, In 2nd
Conference on Pattern Recognition and Image Analysis, vol. 2, pp.
209-216, 2005.
[16] I. Haritaoglu, D. Harwood, and L. Davis,W4S: A real time
system for detecting and tracking people in 2.5D, In Proc. Of the
European Conference on Computer Vision (ECCV), pp. 877-892,
1998.
[17] D. Forsyth and M. Fleck, Body Plans, In Proc. of the
International Conference on Computer Vision and Pattern
Recognition CVPR, pp. 678-683, 1997.

[18] S. Ioffe and D. Forsyth, Probabilistic Methods for Finding

People, International Journal of Computer Vision, vol. 43, no. 1,
pp. 45-68, 2001.
[19] K. Mikolajcyk, C. Schmid, A performance evaluation of
local descriptors, In Proc. Of the International Conference on
Computer Vision & Pattern Recognition (CVPR), pp. 257-263,
2003.
[20] D. Ramanan, D.A. Forsyth, A. Zisserman, Tracking People
by Learning their appearance, IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 29, no. 1, pp. 65-81, 2007.
[21] R. Fergus, P. Perona, and P. Zisserman, Object class
recognition by unsupervised scale invariant learning, In Proc. of
the IEEE Conference on Computer Vision and Pattern
Recognition, vol. 2, pp. 264-271, 2003.
[22] C. Harris and M.J. Stephens, A combined corner and edge
detector, In Fourth Alvey Vision Conference, pp. 147152, 1988.
[23] M. Ester, H. P. Kriegel, J. Sander, X. Xu, A Density-Based
Algorithm for Discovering Clusters in Large Spatial Databases
with Noise, In Proc. Of the 2nd Int. Conf on Knowledge Discovery
and Data Mining (KDD-96), AAAI Press, 1996.
[24] H. Christensen and W. Forstner, Performance
characteristics of vision Algorithms Machine Vision Applications,
vol.9, no. 5-6, pp. 215-218, 1997.