A Multi-configuration Part-based Person Detector
Alvaro Garcia-Martin
1
, Ruben Heras Evangelio
2
and Thomas Sikora
2
1
Video Processing and Understanding Lab, Universidad Autonoma de Madrid, Madrid, Spain
2
Communication Systems Group, Technische Universitat Berlin, Berlin, Germany
Keywords:
People Detection, Part-based Detector, Multi-configuration Body Parts.
Abstract:
People detection is a task that has generated a great interest in the computer vision and specially in the surveil-
lance community. One of the main problems of this task in crowded scenarios is the high number of occlusions
deriving from persons appearing in groups. In this paper, we address this problem by combining individual
body part detectors in a statistical driven way in order to be able to detect persons even in case of failure of any
detection of the body parts, i.e., we propose a generic scheme to deal with partial occlusions. We demonstrate
the validity of our approach and compare it with other state of the art approaches on several public datasets.
In our experiments we consider sequences with different complexities in terms of occupation and therefore
with different number of people present in the scene, in order to highlight the benefits and difficulties of the
approaches considered for evaluation. The results show that our approach improves the results provided by
state of the art approaches specially in the case of crowded scenes.
1 INTRODUCTION
Within the computer vision field, particularly in the
research area of digital image and video processing,
there exists a rich variety of algorithms for segmen-
tation, object detection, event recognition, etc, which
are being used in security systems. The ability to de-
tect people in video and in particular detecting people
in crowded scenarios is the key to a number of multi-
ple applications including video surveillance, group
behavior modeling, crowd disaster prevention, etc.
Due to the rise in popularity of these applications over
the last years, people detection has gradually experi-
enced a great development. In parallel, interest on
reliable strategies to assess the quality of people de-
tection has also grown.
Currently, many different systems exist which try
to solve the problem posed by the task of detecting
people. The state of the art includes several success-
ful solutions working in specific and constrained sce-
narios. However, the detection of people in real world
scenarios such as airports, malls, etc, is still a highly
challenging task due to the multiple appearances that
different persons may have, heavy occlusions, spe-
cially in crowded scenarios, view variations and back-
ground variability.
The work presented in this paper has been focused
on the improvement of people detection in crowded
scenarios. To that aim, we use a part-based person
model and propose a statistical driven way of com-
bining the individual body part detectors in order to
detect persons even in case of failure on the detection
of any of the body parts. Thus, we are able to detect
people with nearly the same reliability whether they
are completely visible (people in front of the group)
or only partially visible (people behind). We validate
our approach and compare with the state of the art
on challenging crowded scenes from multiple public
video datasets.
The remainder of this paper is structured as fol-
lows: Section 2 describes the related state of the art;
Sections 3 describes the proposed approach; Section
4 discusses the experimental results. Finally, Section
5 summarizes the main conclusions and future work.
2 STATE OF THE ART
People detection methods from the state of the art
perform well in scenes with relatively small number
of pedestrians (Doll´ar et al., 2012b; Enzweiler and
Gavrila, 2009; Ger´onimoet al., 2010), but these meth-
ods usually fail or significantly reduce their perfor-
mance in scenes with many subjects that partially oc-
clude each other. Various solutions or improvements
havebeen proposed in order to deal with the occlusion
321
Garcia-Martin A., Heras Evangelio R. and Sikora T..
A Multi-configuration Part-based Person Detector.
DOI: 10.5220/0005126703210328
In Proceedings of the 11th International Conference on Signal Processing and Multimedia Applications (MUSESUAN-2014), pages 321-328
ISBN: 978-989-758-046-8
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
limitations, including directly intrinsic modifications
of the person model or using information from addi-
tional sources.
Most of the methods in the existing literature are
based on the appearance of the object of interest, i.e.,
a person. Appearance based approaches can be classi-
fied attending to the complexity of the model. Simple
person models define the person as a region or shape,
which can be described by means of a holistic model
(Dalal and Triggs, 2005; Doll´ar et al., 2012a). Com-
plex models define the person as collection of multi-
ple regions or shapes, i.e., part-based models, which
can be combined in order to be more flexible regard-
ing different poses and to support partial occlusions
(Felzenszwalb et al., 2010; Leibe et al., 2005). How-
ever, even such approaches have difficulties in dense
environments. The Implicit Shape Model (ISM) of
(Leibe et al., 2005) is improved in (Seemann et al.,
2007) by using a probabilistic formulation in order to
generate a model that is scalable from a general ob-
ject–class detector into a specific object–instance de-
tector, thus making the detection more reliable. The
detector in (Felzenszwalb et al., 2010) is improved
in (Girshick et al., 2011) by using a grammar model
which includes an additional “body part” simulating
possible occlusions. Also based on (Felzenszwalb
et al., 2010), in (Tang et al., 2014) a joint model is
proposed, which is trained to detect single people as
well as pairs of people under varying degrees of oc-
clusion.
Other approaches make use of additional exter-
nal information to the person model in order to in-
crease the detection performance in crowded scenar-
ios. The most typical ones include tracking (Garcia-
Martin and Martinez, 2012), motion (Patzold et al.,
2010), depth or 3D information, etc. The use of per-
son density estimation to improve person localization
and tracking performance in crowded scenes is pro-
posed in (Rodriguez et al., 2011). In (Milan et al.,
2014) a continuous energy minimization framework
for multi-target tracking, which includes explicit oc-
clusion reasoning and appearance modeling, is pre-
sented. Nevertheless, in the work presented in this
paper, we will disregard any possible additional im-
provement which could be achieved by using external
information to the person model. Instead of that, we
concentrate on the person model itself.
Most closely related to our work is the approach in
(Girshick et al., 2011), which demonstratesthe advan-
tages of taking into account in the person model the
possibility of failure or occlusion of some body parts.
In our case, we do not specifically train the model to
capture specific occlusion patterns. We define a more
generic scheme in which the absence of any partic-
ular body part can be modelled by defining multiple
configurations of the part-based models learned dur-
ing the training phase. Therefore, we are able to deal
with occlusions by automatically selecting which of
all the possible person model configurations adjust
better to any kind of occlusion. In particular, we solve
the problem posed to the approach in (Girshick et al.,
2011) by crowded scenarios, where the range of pos-
sible different occlusions is much bigger and, there-
fore, the complexity of the grammar model and its
training increases exponentially.
There are also other approaches that make use of
person models based only on some parts of the body
as the head (Ali and Dailey, 2012) or head and shoul-
ders (Zeng and Ma, 2010) since these are the most
visible parts in crowded scenarios. Our solution can
be considered as a generalization where any possible
body part configuration is evaluated in order to take
advantage not only on this specific simplified models
but also any possible useful configuration.
3 APPROACH
Our proposed approach is based on the detector pre-
sented in (Felzenszwalb et al., 2010) but, instead of
using the confidence provided by each of the individ-
ual body-part detectors for every person candidate, we
define several body-part detectors configurations in
order to robustly cope with partial occlusions, which
profusely appear in crowded scenarios.
3.1 Base Algorithm
The detector in (Felzenszwalb et al., 2010) is a part-
based person model. It consists of mixtures of multi-
scale deformable part models in a star-structure de-
fined by a root model, where the root and each of
the deformable body parts are modeled by a HOG as
firstly proposed in (Dalal and Triggs, 2005).
The detector proposed in (Felzenszwalb et al.,
2010) defines N body parts positioned around the root
filter (n = 0), which models the appearance of the
whole body. The N body parts are computed at twice
the resolution in relation to the root filter in order
to refine the detection based only on the root infor-
mation. Each of the n detectors, included the root
(n = 0, ..., N), is modeled by a 3-tuple (F
n
, v
n,0
, d
n
),
where F
n
is the HOG filter response (detection con-
fidence) for part n; v
n,0
is a two-dimensional vector
defining the relative position of part n with respect
to the anchor position (x
0
, y
0
) of the root; and d
n
is
a four-dimensional vector specifying coefficients of a
quadratic function defining the cost for each possible
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
322
placement of the part relative to the anchor position.
The BP
n
(x, y, s) represents the confidence at pixel po-
sition (x, y) for body part n (n = 0, ..., N) associated
to scale s (s = 1, ..., S). Thus, the confidence score for
part n at scale s is given as
BP
n
(x, y, s) = F
n
(x, y, s) − hd
n
, φ(dx
n
, dy
n
)i (1)
with
(dx
n
, dy
n
) = (x
n
, y
n
) − (2(x
0
, y
0
) + v
n,0
) (2)
givingthe displacement of part n relativeto the anchor
and
φ(dx, dy) =
dx, dy, dx
2
, dy
2
(3)
defining the potential spatial deformation distribu-
tions. Figure 1-(c) shows one example of a multi-part
person model with N = 9.
The final detection confidence or score C(x, y, s) is
computed as the sum of the root and N body parts at
each pixel position and scale.
C(x, y, s) =
N
∑
n=0
BP
n
(x, y, s) (4)
The final multi-scale detection hypotheses are
extracted after a thresholding followed by a non-
maximum suppression process, used to eliminate pos-
sible repeated detections. The chosen threshold or
minimum score required in order to consider the de-
tected object as a person depends directly on the total
number of body parts detections.
3.2 Multiple Person Model
Configurations
The previously described approach (Felzenszwalb
et al., 2010) is based on the detection of several parts
and the combination of all of them. Since the total
score depends tightly on the number of parts detected,
this approach is not able to reliably cope with occlu-
sions. Therefore, it fails to detect people in groups,
where most of the persons are only partially visible.
In order to cope with any kind of body part occlu-
sion, we propose to use multiple person model config-
urations t (t = 1, ..., T) with 1 ≤ T ≤ 2
N
, where each
person model configuration t consist of a subset of M
body parts (m = 1, ..., M), with m ⊂ n of the original
detector (Felzenszwalb et al., 2010) and 1 ≤ M ≤ N.
Thus, the confidence for each configuration is defined
as
C
t
(x, y, s) =
N
∑
n=0
α
t
n
· BP
n
(x, y, s) (5)
where α
t
is a binary selector vector for each con-
figuration t
α
t
n
=
(
1 , n ⊂ t
0 , otherwise
(6)
As in the base algorithm, the final multi-scale de-
tection hypotheses are extracted after a thresholding
followed by a non-maximum suppression process in
order to eliminate possible repeated detections. How-
ever, there are two main differences in our approach
with respect to the base algorithm. In first place, there
is not only one detection threshold, but there is one
for each configuration. Each minimum score required
is chosen to be coherent with the number and kind of
body parts taken into consideration (see Section 3.3).
In second place, we apply the non-maximumsuppres-
sion process to the results provided by all the person
model configurations together.
3.3 Body Parts Contributions
Once defined the different person model configura-
tions, it is necessary to determine the decision thresh-
old or minimum score required for each configura-
tion in relation to the threshold used if considering
the whole set of body parts. To that aim, let con-
sider the confidence or score of each body part n as
a continuous random variable BP
n
and its associated
probability density function f
BP
n
(bp
n
), the final de-
tection confidence as a continuous random variable C
and its associated probability density function f
C
(c).
The minimum confidence k required to consider a de-
tected object as a person corresponds to the probabil-
ity of F
C
(k) = P(C ≤ k).
Analogously, each configuration confidence can
be considered as a continuous random variable
C
t
with an associated probability density function
f
C
t
(c
t
). In order to estimate the minimum confidence
k
t
required for each configuration, it is necessary to
determine a correction factor R
t
that takes into ac-
count the number of body parts included in each con-
figuration and their respective contribution or infor-
mation relevance in relation to the original configu-
ration with N parts. For example, assuming that all
the body parts had the same contribution the correc-
tion factor R
t
=
1
N
could be used for each configura-
tion t. Nevertheless, since the individual part detec-
tors are not equally discriminative, their contribution
to the overall model can not be considered the same.
Therefore, in order to estimate the contribution of
each body part n, we first estimate the similarity of
the distribution of the scores obtained by using the
configuration with the whole set of body parts (F
C
),
with the distribution of the scores obtained by using
AMulti-configurationPart-basedPersonDetector
323
the configuration with all except the considered body
part n. To that aim, we use the Kullback-Leibler Di-
vergence (D
KL
) (Kullback and Leibler, 1951) and de-
fine the similarity KL
n
between each body part BP
n
to
the complete model C as the Kullback-Leibler Diver-
gence between the distribution F
C
and the distribution
without that body part n, F
C
′
:
KL
n
= D
KL
(F
C
||F
C
′
), beingC
′
=
N
∑
i=0,i6=n
BP
i
(7)
This measure is normalized
¯
KL
n
so that
∑
N
n=1
¯
KL
n
= 1. Finally, the correction factor R
t
is computed as the accumulative body parts contribu-
tions:
R
t
=
N
∑
n=0
α
t
n
·
¯
KL
n
(8)
A factor of R= 1 means that there is not necessary
any correction on the decision threshold because the
considered configuration corresponds to the use of all
the body parts.
The minimum confidence k
t
required for each
configuration t with associated probability F
C
t
(k
t
) =
P(C
t
≤ k
t
) is modified according to the original per-
son model confidence k and the corresponding correc-
tion factor R
t
:
F
C
t
(k
t
) = 1− R
t
(1− F
C
(k)) (9)
Therefore, the final probability F
C
t
(k
t
) required
for each configuration is defined between the original
F
C
(k) and 1 (F
C
(k) ≤ F
C
t
(k
t
) ≤ 1). The simpler the
person model (less body parts), the higher the prob-
ability (i.e., the confidence) required to detect a per-
son and vice versa. Figure 1 shows examples of per-
son model configurations, distributions, correspond-
ing correction factors and minimum confidence re-
quired k
t
.
4 EXPERIMENTAL RESULTS
In order to evaluate our people detection approach,
we have tested it across several publicly available
datasets and compare its results with those pro-
vided by the base algorithm DTDP (Discrimina-
tively Trained Deformable Parts (Felzenszwalb et al.,
2010)), with those provided by the ISM (Implicit
Shape Model (Leibe et al., 2005)) and those provided
by one of the most recent and cited people detection
approaches from the state of the art ACF (Aggregate
Channel Features (Doll´ar et al., 2012a)). While the
DTDP and ISM are part-based detectors, the ACF is
a holistic approach.
As presented in this paper, our proposed approach
consists of multiple person model configurations. In
particular, according to the person model of nine body
parts (see Figure 1), we define twenty different con-
figurations (T = 20). Every configuration includes at
least the root and head body parts, from this basic
configuration (t = 1), we add progressively configu-
rations with consecutive body parts, i.e., root-head-
left shoulder, root-head-left and right shoulder, root-
head-left shoulder and left trunk, etc. The last config-
uration (t = 20) includes all the body parts and cor-
responds to the original configuration of the base al-
gorithm DTDP (Felzenszwalb et al., 2010). The body
part contributions (see section 3.3) have been trained
using the code provided in (Girshick et al., ) and the
INRIA dataset (Dalal and Triggs, 2005). The DTDP
results have been obtained using the available code
(Girshick et al., ) and the ACF results have been ob-
tained using the available code (Doll´ar et al., 2012a).
We evaluate all three methods on eleven challeng-
ing, publicly available video sequences with ground
truth (Milan et al., 2014) (TUD-Stadtmitte (Andriluka
et al., 2010), TUD-Campus and TUD-Crossing (An-
driluka et al., 2008), S1L1 (1 and 2), S1L2 (1 and
2), S2L1, S2L2, S2L3 and S3L1). The first three se-
quences are recorded in real-world busy streets, the
complexity in terms of crowdor occlusions is medium
or low (less than 10 pedestrians are present simul-
taneously). The last eight sequences are part of the
PETS 2009/2010 benchmark (PETS, ). We only use
the first view of each sequence in all our experiments.
They are recorded outdoors from an elevated view-
point, corresponding to a typical surveillance setup.
The sequences are classified originally according to
three scenarios (S1, S2 and S3) and three progressive
difficulty levels (L1, L2 and L3) for each scenario.
These scenarios include higher complexity in terms of
crowds and occlusions than the previous ones (gener-
ally more than 10 pedestrians are present simultane-
ously).
We classify the whole set of sequences indepen-
dently of the original scenario purpose (TUD se-
quences for people detection, S1 for person count
and density estimation, S2 for people tracking and
S3 for flow analysis and event recognition). In our
experiments, we classify the sequences according to
the number of people present simultaneously and,
therefore, the degree of occupation of the scene (low,
medium or high). Table 1 includes a description
of each sequence in terms of occupation (number
of pedestrian present simultaneously) and complex-
ity classification. Figure 2 shows sample images of
the used sequences.
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
324
Example M
1
= 2 Example M
2
= 6 Example M
3
= N = 9
0 ≤ R
1
≤ R
2
R
1
≤ R
2
≤ R
3
R
3
= 1
F
C
2
(k
2
) ≤ F
C
1
(k
1
) ≤ 1 F
C
3
(k
3
) ≤ F
C
2
(k
2
) ≤ F
C
1
(k
1
) F
C
3
(k
3
) = F
C
(k)
(a) (b) (c)
Figure 1: Examples of person model configurations, distributions, corresponding correction factors and minimum confidence
required. (a) Example with root and head. (b) Example with root and 5 body parts. (c) Example with root and 8 body parts
(original model with N = 9 (Felzenszwalb et al., 2010)).
(a) (b)
(c) (d)
Figure 2: Experimental sequences examples: (a) TUD-Crossing, (b) PETS2009-S2-L1, (c) PETS2009-S1-L1-2 and (d)
PETS2009-S1-L2-1.
AMulti-configurationPart-basedPersonDetector
325
(a) (b)
(c) (d)
Figure 3:
Examples of people detection performance in terms of Precision-Recall curves.
Table 1: Experimental results. Occupation in terms of number of pedestrians present simultaneously. Complexity classifica-
tion. People detection performance in terms of
area under the Precision-Recall curve (AUC-PR). Percentage increase (%∆
1
and %∆
2
) calculated with respect to original performance DTDP and ACF respectively.
Occupation
Complexity
AUR-PR
Up to # DTDP ISM ACF Ours %∆
1
%∆
2
TUD-Campus 8 Low 0.76 0.76 0.80 0.80 +5.3 0.0
TUD-Stadmitte 6 Low 0.79 0.71 0.83 0.81 +2.5 -2.4
TUD-Crossing 8 Low 0.85 0.84 0.88 0.87 +2.4 -1.1
PETS2009-S1L1-1 34 Medium 0.63 0.45 0.63 0.67 +6.3 +6.3
PETS2009-S1L1-2 26 Medium 0.73 0.49 0.68 0.80 +9.6 +17.6
PETS2009-S1L2-1 42 High 0.48 0.30 0.44 0.56 +16.7 +27.3
PETS2009-S1L2-2 40 High 0.50 0.36 0.51 0.57 +14.0 +11.8
PETS2009-S2L1 8 Low 0.93 0.78 0.85 0.95 +2.2 +11.8
PETS2009-S2L2 35 Medium 0.66 0.55 0.58 0.75 +13.6 +29.3
PETS2009-S2L3 42 High 0.55 0.34 0.47 0.62 +12.7 +31.9
PETS2009-S3L1 7 Low 0.93 0.82 0.94 0.95 +2.2 +1.1
For evaluating people detection performance
based on ground-truth, we aim to compare the overall
performance of different detection systems, so we
have chosen the Precision-Recall (PR) evaluation
method. In order to evaluate not only the (binary)
yes/no detection but also the precise pedestrians
locations and extents, we use also the three criteria,
defined by (Leibe et al., 2005), that allow comparing
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
326
hypotheses at different scales: the relative distance,
cover, and overlap. Only one hypothesis per object
is accepted as correct, so any additional hypothesis
on the same object is considered as a false positive.
In addition, we use the integrated Average Precision
(AP) to summarize the overall performance, repre-
sented geometrically as the area under the PR curve
(AUC-PR).
Figure 3 shows examples of results in terms
of PR curves over the sequences TUD-Crossing,
PETS2009-S2-L1, PETS2009-S1-L1-2 and
PETS2009-S1-L2-1. Table 1 shows results in
terms of AUC-PR. In both cases, the results show
clearly how the performance decreases for every
tested approach from the simplest sequences (TUD
sequences) to the medium and high complexity
sequences (PETS2009). The ACF detector provides
the best results over the simplest scenarios, where
there are few occlusions, but it provides worse
results than the DTDP over the scenarios with more
occlusions. The main reason for this behavior is that
the ACF detector is based on a holistic person model
and presents difficulties dealing with occlusions.
The DTDP detector provides good results over
the simplest sequences but worse than the ACF
detector; however, the DTDP detector is based on a
part-based person model and for this reason provides
better results over the complex sequences with more
occlusions. The ISM detector provides similar results
than the DTDP detector over the simplest sequences
but the worse results of all the three detectors over
the complex sequences. In this case, the feature-part-
based model is not robust enough to deal with partial
occlusions.
Our proposal, the “DTDP multi configurations”
provides better results than the DTDP detector in all
the cases. It is clear that the improvement is more
significant in those scenarios with higher complexity
or occupation (PETS2009 sequences with high
occupation, 12.7-16.7% improvement respect to the
original DTDP detector) than in those scenarios with
lower complexity (TUD and PETS sequences with
low occupation, 2.4-5.3% improvement). This was
expected, since the improvement possibilities on
those sequences with more occlusion difficulties is
higher.
Comparing our approach with the ACF detector,
our proposal provides similar or slightly worse
results on simple scenarios (TUD sequences with low
occupation: -2.4 to 0% improvement with respect
to the ACF detector), due to the lower performance
of the base detector DTDP. However, over those
scenarios with higher complexity or occupation,
our approach provides significant improvements
(PETS2009 sequences with high occupation, 11.8-
31.9% improvement respect to the original DTDP
detector).
The tests have been performed on aN AMD
Opteron(tm) Processor 4386 with a 4xCPU fre-
quency of 3 GHz and 4GB RAM. During our
experiments, the original DTDP computational
cost is around 3.6 seconds per frame with 640x480
images (TUD sequences) and around 4.9 seconds
per frame with 768x576 images (PETS sequences).
The proposed
multi-configurations approach includes
the same main core of the original approach and the
additional computational cost of computing T config-
uration confidences (see equation 5) instead of only
one (see equation 4). In the case of twenty different
configurations (T = 20), the computational cost
is
around 4.9 seconds per frame with 640x480 images
(TUD sequences) and around 6.2 seconds per frame
with 768x576 images (PETS sequences). Assuming
that we are able to run every configuration in parallel,
the final computational cost will be established by
the configuration with the maximum number of body
parts, i.e., the original DTDP computational cost.
To sum up, the results show how the proposed
combination of multiple configurationsis more robust
to partial occlusions than the original DTDP detector.
In particular, our proposed detector provides better
results than all other reference detectors from the
state of the art (DTDP, ISM and ACF) over typical
sequences with a high degree of occupation (groups
of people together), where the presence of occlusions
is typical.
5 CONCLUSIONS
People detection methods from the state of the art per-
form well in scenes with relatively few people, but are
severely challenged by scenes with many subjects that
partially occlude each other. We observe that typical
occlusions are due to overlaps between people and
propose a people detector tailored to various occlu-
sion levels. We propose a generic multiple body parts
combination framework in order to deal with these
specific partial occlusions in crowded scenarios.
We have validated our approach and compared it
with other state of the art approaches on several pub-
lic datasets. The results, over sequences with differ-
ent number of people present in the scene simulta-
neously, demonstrate the achieved improvements in
typical crowded scenes where the number and range
of possible different occlusions are much higher than
in simpler scenarios.
As future work, we will try to extrapolate this
AMulti-configurationPart-basedPersonDetector
327
scheme to other people detectors or even to any kind
of object part-based approach. In addition, we pro-
pose to explore other methods in order to estimate
both the probability density functions and the simi-
larity between them.
ACKNOWLEDGEMENTS
This work has been done while visiting the Commu-
nication Systems Group at the Technische Universit¨at
Berlin (Germany) under the supervision of Prof. Dr.-
Ing. Thomas Sikora. This work has been partially
supported by the Universidad Aut´onoma de Madrid
(“Programa propio de ayudas para estancias breves
en Espa˜na y extranjero para Personal Docente e In-
vestigador en Formaci´on de la UAM”), by the Span-
ish Government (TEC2011-25995 EventVideo) and
by the European Community’s FP7 under grant agree-
ment number 261776 (MOSAIC).
REFERENCES
Ali, I. and Dailey, M. N. (2012). Multiple human tracking
in high-density crowds. Image and Vision Computing,
30(12):966 – 977.
Andriluka, M., Roth, S., and Schiele, B. (2008).
People-tracking-by-detection and people-detection-
by-tracking. In Proc. of CVPR, pages 1–8.
Andriluka, M., Roth, S., and Schiele, B. (2010). Monocular
3d pose estimation and tracking by detection. In Proc.
of CVPR, pages 623–630.
Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-
dients for human detection. In Proc. of CVPR, pages
886–893.
Doll´ar, P., Appel, R., and Kienzle, W. (2012a). Crosstalk
cascades for frame-rate pedestrian detection. In Proc.
of ECCV, number 645-659.
Doll´ar, P., Wojek, C., Schiele, B., and Perona, P. (2012b).
Pedestrian detection: An evaluation of the state of the
art. IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence, 34(4):743–761.
Enzweiler, M. and Gavrila, D. M. (2009). Monocular pedes-
trian detection: Survey and experiments. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence,
31(12):2179–2195.
Felzenszwalb, P. F., Girshick, R. B., McAllester, D., and
Ramanan, D. (2010). Object detection with discrim-
inatively trained part-based models. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence,
32(9):1627–1645.
Garcia-Martin, A. and Martinez, J. M. (2012). On collab-
orative people detection and tracking in complex sce-
narios. Image and Vision Computing, 30(4):345–354.
Ger´onimo, D., L´opez, A. M., Sappa, A. D., and Graf, T.
(2010). Survey of pedestrian detection for advanced
driver assistance systems. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 32(7):1239–
1258.
Girshick, R. B., Felzenszwalb, P. F., and McAllester, D.
Discriminatively trained deformable part models, re-
lease 4. http://people.cs.uchicago.edu/ rbg/latent-
release4/.
Girshick, R. B., Felzenszwalb, P. F., and Mcallester, D.
(2011). Object detection with grammar models. In
Proc. of NIPS.
Kullback, S. and Leibler, R. A. (1951). On information
and sufficiency. The Annals of Mathematical Statis-
tics, 22(1):79–86.
Leibe, B., Seemann, E., and Schiele, B. (2005). Pedestrian
detection in crowded scenes. In Proc. of CVPR, pages
878–885.
Milan, A., Roth, S., and Schindler, K. (2014). Continuous
energy minimization for multitarget tracking. IEEE
Transactions on Pattern Analysis and Machine Intel-
ligence, 36(1):58–72.
Patzold, M., Evangelio, R. H., and Sikora, T. (2010). Count-
ing people in crowded environments by fusion of
shape and motion information. In Proceedings of
the 2010 7th IEEE International Conference on Ad-
vanced Video and Signal Based Surveillance, AVSS
’10, pages 157–164, Washington, DC, USA. IEEE
Computer Society.
PETS. International workshop on performance
evaluation of tracking and surveillance,
http://www.cvg.rdg.ac.uk/pets2009/index.html.
Rodriguez, M., Laptev, I., Sivic, J., and Audibert, J.-Y.
(2011). Density-aware person detection and tracking
in crowds. In Proc. of ICCV, pages 2423–2430.
Seemann, E., Fritz, M., and Schiele, B. (2007). To-
wards robust pedestrian detection in crowded image
sequences. In Proc. of CVPR, pages 1–8.
Tang, S., Andriluka, M., and Schiele, B. (2014). Detection
and tracking of occluded people. International Jour-
nal of Computer Vision.
Zeng, C. and Ma, H. (2010). Robust head-shoulder detec-
tion by pca-based multilevel hog-lbp detector for peo-
ple counting. In Proc. of ICPR, pages 2069–2072.
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
328
