Launch These Manhunts! Shaping the Synergy Maps for Multi-camera

Detection

Muhammad Owais Mehmood

, Sebastien Ambellouis

and Catherine Achard

LEOST, French Institute of Science and Technology for Transport, Spatial Planning, Development and Networks,

Villeneuve D’Ascq, France

Sorbonne Universites, UPMC Univ. Paris 06 and CNRS UMR 7222, ISIR, F-75005, Paris, France

Keywords:

People Localization, Ghost Pruning, Multi-camera Surveillance, Shape Representations, Pattern Recognition.

Abstract:

We present a method for multi-camera people detection based on the multi-view geometry. We propose to

create a synergy map by the projection of foreground masks across all camera views on the ground plane and

the planes parallel to the ground. This leads to signiﬁcant values on locations where people are present, and

also to a particular shape around these values. Moreover, a well-known ghost phenomena appears i.e. when

these shapes corresponding to different persons are fused then the false detections are also generated. In this

article, the ﬁrst improvement is the robust detection of the candidate detection locations, namely keypoints,

from the synergy map based on a watershed transform. Then, in order to reduce the false positives, mainly due

to the ghost phenomena, we check if the particular shape, for an ideal person, is present or not. This shape,

that is different for each location of the synergy map, is generated for each keypoint, assuming the presence

of a person, and with the knowledge of the scene geometry. Finally, the real shape and the synthetic one are

compared using a similarity measure that is similar to correlation. Another improvement proposed in this article

is the use of unsupervised clustering, performed on the measures obtained at all the keypoints. It allows to

automatically ﬁnd the optimal threshold on the measure, and thus to decide about people detection. We have

compared our method to the recent state-of-the-art techniques on a publicly available dataset and have shown

that it reduces the detection errors.

1 INTRODUCTION

People detection is a well-studied issue in computer vi-

sion with applications such as in the video surveillance

systems. The challenges pertaining to people detection

include the involvement of human articulations, scale

and appearance based variations, occlusion, density,

and environment clutter. Extensive research has been

performed on the single camera human detection al-

gorithms; however, these systems remain limited in

their ability to handle occlusions, dense and cluttered

environments (Doll

ar et al., 2012).

Recently, researchers have focused on the multi-

view algorithms as a possible solution to overcome the

limitations of the single camera detection techniques.

Khan and Shah (Khan and Shah, 2009) present a ho-

mographic occupancy constraint and apply it across

multiple planes and camera views to obtain wrapped

foreground occupancies. They further introduce syn-

ergy map which is the fusion of the wrapped fore-

ground occupancies. This synergy map, however, suf-

fers from several false positives called ghosts. Eshel

and Moses (Eshel and Moses, 2010) perform people

detection and tracking in multiple camera systems in-

stalled at the top/head level elevations. The head de-

tection occurs by applying intensity correlation across

the head level planes that are aligned by the planar

homography constraint. This method demonstrates

increased performances but remains limited to head

level conﬁguration.

Besides the two earlier geometric methods, the

method in (Fleuret et al., 2008) deﬁnes a Probabilistic

Occupancy Map (POM) that performs people detection

assuming rectangles of average human size placed at

a discretized ground plane. However, this method

suffers from high false positive rate and has a high

computational cost. In (Utasi and Benedek, 2011;

Utasi and Benedek, 2013), the authors improve the

localization accuracy by performing an optimization

process that ﬁts a cylinder, modeling a person, on the

multi-planar features.

All previous methods suffer from ghosts but pro-

528

Owais Mehmood M., Ambellouis S. and Achard C..

Launch These Manhunts! Shaping the Synergy Maps for Multi-camera Detection.

DOI: 10.5220/0005355805280535

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 528-535

ISBN: 978-989-758-090-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

(a) Camera View 2 (b) Synergy Map

Figure 1: Synergy map generated for an area monitored in the frame 593 of PETS 2009 dataset. Brighter red indicates higher

detection probability. The positive green symbols indicate the positions of the three persons (white circles in the camera view).

The yellow ellipses correspond to the false detections. The outer rectangle represents the monitored area, and rectangles around

the people are obtained from the ground truth.

vide no explicit solution for it. For the geometric,

multi-planar or homographic techniques, fusion of the

projections corresponding to different people in the

synergy map could generate false detections, these de-

tections of non-corresponding regions in the projected

space are referred to as the ghosts. Color features have

been used to remove the incorrect correspondences

across the two camera views (Ren et al., 2012). These

features require a certain degree of color consistency

across the views. Similar to color features, background

occupancy map has been proposed in (Mehmood et al.,

2014), but the method is not robust to signiﬁcant per-

spective effects. Probability of the presence of a ghost

based on the relative position of the objects from the

cameras has been explored in (Evans et al., 2012). This

method requires temporal information in order to re-

move the false detections.

We propose a novel method for people detection

across multiple synchronized views based on coher-

ence analysis. We have observed that it is possible

to generate a shape model in the synergy map based

on the location of an object in the scene. We refer to

this shape as the Synergy Shape Model (SSM). This

shape model is a map created by modeling the person

as the axis of a cylinder, at a given 3D location, fol-

lowed by the fusion of the multiplanar projections of

its synthetic images. We apply this model for ghost

pruning using a similarity measure between the SSM

and the real synergy map. The 3D locations, at which

our model is processed, are obtained by the application

of local maxima and a modiﬁed watershed transform

on the real synergy map. Thus, our algorithm is based

on the knowledge of the multi-view scene geometry.

Finally, we perform cluster analysis on the similarity

measures to automatically deﬁne the decision bound-

ary for people detection.

Our contributions include the introduction of a

technique for robust candidate detection in synergy

map, the generation of its corresponding shape model

using the knowledge of the scene geometry, and the

use of unsupervised clustering for automated optimal

threshold calculation over shape similarity measures

for multi-view people detection. Compared to the

state-of-the-art, our method, for ghost pruning, utilizes

all the camera views, including those with signiﬁcant

perspective effects ignored in (Mehmood et al., 2014),

in order to achieve multi-view reasoning. The pro-

posed algorithm does not require the use of color (Ren

et al., 2012) or temporal information (Evans et al.,

2012) but analyzes the multi-view geometry. More-

over, our proposed algorithm is not limited to only one

top level camera conﬁguration, as in (Eshel and Moses,

2010). Finally, we propose a quantitative analysis with

the state-of-the-art (Fleuret et al., 2008; Mehmood

et al., 2014; Utasi and Benedek, 2013) to demonstrate

the efﬁciency of our technique on a popular public

dataset (PETS, 2009). Compared to the recent state-of-

the-art (Utasi and Benedek, 2011; Utasi and Benedek,

2013), our method also requires less input parameters.

The rest of the paper is structured as follows. The

proposed approach is presented in Section 2. In Sec-

tion 3, we present the evaluation of our method, the

quantitative analysis and insights into the efﬁciency of

our approach. The paper concludes in Section 4.

2 PROPOSED METHOD

The proposed method makes use of the foreground

masks obtained for each image of the calibrated multi-

view systems. In the present implementation, we use

the background subtraction method proposed in (Yao

LaunchTheseManhuntsShapingtheSynergyMapsforMulti-cameraDetection

529

(a) Camera View 1 (b) Synergy Map

Figure 2: Our proposed method generates (b) a synergy map, and (d) a synthetic Synergy Shape Model against the multi-view

observations. The person visible in the Area of Interest (AOI) is modeled by the longitudinal axis of the cylinder (red). The

height of this axis is denoted by L.

and Odobez, 2007), and the camera calibration tech-

nique proposed by Tsai (Tsai, 1992). The algorithm

proceeds with the generation of foreground silhou-

ettes in each camera view. The silhouettes are pro-

jected across multiple planes parallel to the ground

plane, with heights between zero and the average hu-

man height. These projections are then merged to

produce the synergy map that has signiﬁcant values at

the locations corresponding to people or ghost. As the

local shape of the synergy map is not the same in these

two cases, therefore the synthetic synergy maps corre-

sponding to an ideal person is generated and compared

to the real synergy maps using a similarity measure.

Finally, we perform cluster analysis of these similarity

measures to differentiate between the human and the

ghosts.

2.1 Multi-planar Projections & Synergy

Map

Foreground silhouette maps are projected on the planes

from the ground i.e.

to the plane

parallel to the

ground plane at height z, as shown in Fig. 3. As we

don’t have the real height of the people therefore z

takes values in the range of the typical human heights.

The projection on the ground plane i.e.

is obtained

using the camera calibration (Tsai, 1992). The projec-

tions on the parallel planes can be efﬁciently computed

from P

by using the following equations:

= x

− (x

− x

)z/h

, (1)

= y

− (y

− y

)z/h

, (2)

where

)

represents the arbitrary position of a

camera located at height

)

is the projection of

a point projected to the plane

and whose coordinates

are

)

(Utasi and Benedek, 2013). In case

the calibration information is not available then the

multi-planar projections can be generated using the

homography based technique as proposed in (Khan

and Shah, 2009).

Following this, a synergy map is computed by sum-

ming all the planes, and all the camera views. Fig. 1

shows an illustration of a synergy map which was

generated using three camera views and 211 differ-

ent planes between 0 and 210 cm. It is possible that

the foreground masks of a particular camera may be

missing information, for example, due to occlusion,

clutter, or inconsistencies in the generation of the fore-

grounds (Cristani et al., 2010). The fusion of the in-

formation present across all the camera views and in-

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

530

cluding multiple heights introduces robustness to such

corruption, which may arise in an individual camera

view or at a particular planar height. As speciﬁed be-

fore, signiﬁcant values in the synergy map correspond

to the people or ghosts. These values are produced by

a chance alignment of projections coming from sev-

eral people, as it can be seen in Fig. 1. Our algorithm

proceeds by extracting the locations of the synergy

map with the signiﬁcant values namely the keypoint

extraction.

2.2 Keypoint Extraction

Keypoints are extracted as the local maxima in the

synergy map

calculated using 8

8 pixel blocks. In

order to reduce multiple detections due to intensity

variations and noise, we use a modiﬁed version of

the watershed transform with markers (Beucher and

Meyer, 1993). This keeps the pedestrian population

tractable and results in a compact description of the

global scenario that can be further analyzed.

We sort the local maxima in descending order and

invert them to deﬁne our markers. The general water-

shed algorithm deﬁnes catchment basins or watershed

regions, which are separated by the watershed lines.

The watershed lines in our case are deﬁned relative

to the local maxima, that is, we introduce a tolerance

threshold

such that a local maxima is a keypoint if

and only if the pixels in the catchment basin — pixels

greater than the difference of the maxima and toler-

ance — around the local maxima are less than itself.

This can also be understood as that the local maximum

is accepted only if its topographical prominence is

greater than the tolerance threshold. Further, if there

exist multiple similar local maxima in one catchment

basin then we deﬁne the keypoint at their geometric

center. We have shown a 1D illustration of this process

in Fig. 4. Finally, we obtain a set

P = {p

},n = 1 ...N

of N keypoints that can correspond to people or ghost.

)

Figure 3: Projection of the cylindrical axis of height L

(green), corresponding to a person, to the ground (blue)

and a parallel plane at height z (red).

local maxima

(a)

watershed regions

crest

x x x

(b)

Figure 4: One-dimensional illustration of the keypoint ex-

traction based on the watershed transform applied to the

local maxima. (a) Local maxima are extracted on an arbi-

trary distance ﬁeld D. (b) Inverted local maxima are treated

as markers on which the marker based watershed transform

is applied. The crest here acts as the watershed line and

is deﬁned by the

parameter. The local maxima must be

the greatest value in its region. In case of multiple similar

local maxima, the geometric center is taken as the keypoint

location.

2.3 Ghost Pruning

Some false positives or ghosts appear in the set of key-

points obtained (see Fig. 1). Several methods have

already been proposed in the literature to remove them.

Evans et al. (Evans et al., 2012) found that the ghost

detections are probable along the lines from the camera

center to the center of the objects of interest, intuited

as “star” shape at the object, having “streaked legs”

corresponding to the lines (see Fig. 2(d)). In this work,

we propose a novel model that plays a role in ghost

pruning using the shape cues deﬁned around these

“star” shapes and “streaked legs”. Liu et al. (Liu et al.,

2013), for robust auto-calibration, models the pedes-

trian blobs using two end points of the axes of the

ellipses, represented by the vanishing point and esti-

mating the 3D blob heights resembling the real world

distribution of human heights.

Following this, we deﬁne a shape for each person,

represented by the longitudinal axis of a cylinder in the

3D coordinate system of the scene. Let us assume that

the people are standing on a ﬂat ground. We monitor

a rectangular Area of Interest (AOI) in the

ground

plane, and we attempt to model the shape of each

possible pedestrian in

. Thus, the free parameters

of the given longitudinal axis of the cylinder are its

p = (x,y)

coordinate in the ground plane and the length

L. This is illustrated in Fig. 2.

We employ a discrete space of objects in the ground

plane of the AOI, consisting of

× S

locations. For

each keypoint

detected in this space, synthesized

LaunchTheseManhuntsShapingtheSynergyMapsforMulti-cameraDetection

531

(a) Person (b) Person

Figure 5: Illustration of the proposed method on frame 593

of the PETS 2009 dataset. The Synergy Shape Model is

represented by the green colour whereas the synergy maps

are red. It can be observed that the overlap between two is

higher for the persons (a),(b), versus the ghosts (c),(d).

camera view

n,c

, corresponding to its approximated

cylindrical axis, are generated using the camera cal-

ibration matrices i.e. in this case for the cth camera.

Let

denote the set of synthetic images created for

the keypoint p

= {I

n,c

}, c = 1 . . .C. (3)

In this paper, we use the fact that a person generates

not only a signiﬁcant value in the synergy map but also

a particular shape around this point — a shape that is

not present in case of a ghost (see Fig. 5). This shape

is given by the geometry of the cameras and is always

the same for a point in the ground plane. Thus, a way

to decide if the keypoint is a people or a ghost is to

analyze this shape. Therefore, for each keypoint

we generate a synthetic synergy map

, the SSM,

that corresponds to the synergy map to be observed

if a person is present at the location

: the person

is modelled by a vertical line, the axis of a cylinder,

with height L (see Fig. 2). The synthetic images

corresponding to this line are then fused to create the

synthetic synergy map. By comparing

to the real

synergy map

, we can conclude on the similarity

between the shapes and thus, on the detection:

D(R ,S

) =

∑

[min(R (p),S

(p))], (4)

here

is the sum of

(p)

over all pixel locations

of the synergy map. This can be understood as conﬁ-

dence of the hypothesis that a person is present at the

keypoint p

2.4 People Detection

Given the similarity measures distribution over all

dataset, we can proceed towards univariate cluster

analysis to group the similarity measures into class

intervals corresponding to people and ghosts. Due

to cluster analysis, the decision threshold for a key-

point

, to be a person or a ghost, is automatically

computed. Most clustering or vector quantization algo-

rithms can be classiﬁed into partitional or hierarchical

algorithms. Hierarchical clustering algorithms do not

require pre-speciﬁcation of the number of clusters, are

primarily deterministic, but computationally expen-

sive. Partitional or ﬂat clustering algorithms deﬁne a

set of disjoint clusters and are suited for large datasets

where computational efﬁciency is important. How-

ever, as no consensus is present on this issue (Manning

et al., 2008), therefore, we use both partitional and

hierarchical methods.

For hierarchical clustering, we use the Unweighted

Pair Group Method with Arithmetic Mean (UPGMA)

agglomerative clustering method (Sokal and Michener,

1958). We use Euclidean distance for the generation

of the distance matrix. Hierarchical clustering is more

suited to our univariate data because it doesn’t have

enough structure, relative to multi-dimensional data,

and the computational costs are not important. We se-

lect UPGMA clustering because it provides a suitable

trade-off between the complete-link method’s sensitiv-

ity to outliers and that of single-link to form dendo-

gram chains longer than the intuitive notion of com-

pact, spherical clusters. The work in (Li et al., 2008)

also selects agglomerative clustering technique for hu-

man detection in 3D space.

For partitional clustering, we use univariate Kernel

Density Estimation with Epanechnikov kernel (Scott,

1992) and Mixture of Gaussians Expectation Maxi-

mization (MoG-EM) method (univariate, unequal vari-

ance) (Fraley and Raftery, 2007). For KDE, we use

the local minimum to separate the clusters. In case

of MoG-EM algorithm, we assume as a priori that

the number of real objects exceed ghosts, hence we

Figure 6: Estimated locations of the people localized, repre-

sented by the white circles surrounded by the ground truth

rectangles.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

532

Table 1: Comparison of proposed SSM with other techniques using the best parameters to minimize TER. Parameters:

= 31

pixels, L = 175 cm.

Method TER FDR MDR MIR

POM (Fleuret et al., 2008) 0.252 0.179 0.073 0.000

BOM (Mehmood et al., 2014) 0.123 0.029 0.094 0.000

3DMPP (Utasi and Benedek, 2013) 0.122 0.020 0.096 0.006

Prop. SSM 0.076 0.025 0.051 0.000

can have a distribution for ghosts with lower variance,

centred around a mean corresponding to the lower

measure of similarity.

In terms of parameters, hierarchical clustering is

the most suited because KDE requires a bandwidth

speciﬁcation, while a prior has been deﬁned in the

case of MoG-EM method. For all three methods, we

know that the data must be divided into two classes or

clusters. We present the three algorithms as a more fair

means of demonstrating the application of clustering

on our univariate data.

3 EXPERIMENTS

We have compared our method to the state-of-the-art

techniques: POM (Fleuret et al., 2008), 3D Marked

Point Process (3DMPP) model (Utasi and Benedek,

2013), and Back-projected Occupancy Map (BOM)

(Mehmood et al., 2014). For the evaluation of these

methods we have used a subset of City center se-

quence from PETS 2009 dataset (PETS, 2009) as de-

ﬁned in (Utasi and Benedek, 2011; Utasi and Benedek,

2013). The evaluation sequence contains 400 outdoor

scene images obtained from three camera views. The

dataset deﬁnes an overlapping AOI of size 12.2 m

14.9 m, for which the ground truth annotations are

(a) False Detection (b) Missed Detection

Figure 7: Examples of detection errors. (a) false detection

(white cross), (b) missed detection (white rectangle), and (c)

multiple detections (white circle and positive symbol).

also provided. Camera calibration and time synchro-

nization errors are present in the dataset (PETS, 2009;

Utasi and Benedek, 2013) which explains the non-

convergence of the three lines to one point, for exam-

ple, in Fig. 2(d). The maximum number of people

simultaneously monitored in the AOI is 8. For POM,

we report the best results from (Utasi and Benedek,

2013).

For numerical comparison, we use the projected

position error metrics as deﬁned in (Utasi and Benedek,

2013). False detections (FDs) is the count of de-

tections not corresponding to the ground truth (see

Fig. 7(a)). Missed detections (MDs) is the count of

ground truths not detected (see Fig. 7(b)). Multiple

instances (MIs) is the count of multiple estimates as-

signed to the ground truths (see Fig. 7(c)). Total error

(TE) is the sum of FDs, MDs, and MIs.

FDs, MDs, MIs, and TE are expressed in the per-

cent of the number of objects thus the false detections

rate (FDR), missed detections rate (MDR), multiple

instances rate (MIR), and the total error rate (TER).

Here, MDR

≤

1 and MIR

≤ 1

, but FDR and hence

TER may exceed 1.

Proposed Method.

The foreground masks are gener-

ated using the default parameters as deﬁned in (Yao

and Odobez, 2007). For visualization, the results are

back projected to all of the camera views, for exam-

ple, the ﬁrst camera view in Fig. 6. Multi-planar pro-

jections are generated at a constant 2 cm resolution.

For similarity with (Fleuret et al., 2008; Utasi and

Benedek, 2011; Utasi and Benedek, 2013), we ﬁx

the height L to 175 cm, and the SSM is generated

for 56 different planes between 155 and 210 cm (see

Fig. 2(d)). The proposed method has two main pa-

rameters: tolerance threshold

and height L of the

cylindrical axis. Compared to the evaluation of sev-

eral parameters in (Utasi and Benedek, 2011; Utasi

and Benedek, 2013), we have reduced our algorithm

to only one parameter. Therefore, our evaluation of

the proposed algorithm is limited to

. For the cluster

analysis, hierarchical clustering is sufﬁcient (Li et al.,

2008) but we also present the results of other methods.

Similarly, we also show the effect of assigning a value

other than 175 cm to L.

Quantitative Comparison.

We report the evaluation

results for our algorithm in Tab. 1. Considering TER,

LaunchTheseManhuntsShapingtheSynergyMapsforMulti-cameraDetection

533

20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

τ (pixels)

TER (%)

L=165 cm

L=175 cm

L=185 cm

(a)

40 50 60 70 80 90 100

100

Recall

Precision

L=165 cm

L=175 cm

L=185 cm

τ=31 pixels

(b)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.2

0.4

0.6

0.8

False positive rate

True positive rate

L=165 cm

L=175 cm

L=185 cm

τ=31 pixels

(c)

Figure 8: Evaluation of the proposed Synergy Shape Model with different parameter settings. (a) Total Error Rate (TER) as a

function of the

and L parameters. (b) Precision/Recall curves and (c) Receiver Operating Characteristic (ROC) curves in

function of L.

Table 2: Comparison of the different clustering techniques.

The best parameters for minimum TER are used. Parameters:

= 31 pixels, L = 175 cm, and for KDE, the bandwidth of

the kernel is 0.04.

Method TER FDR MDR MIR

KDE 0.077 0.026 0.051 0.000

MoG-EM 0.076 0.025 0.051 0.000

UPGMA 0.076 0.025 0.051 0.000

we observe a 4.6% improvement versus 3DMPP, 4.7%

versus BOM, and 12.9% versus POM. It can be ob-

served from Tab. 2 that the results remain consistent

in spite of the clustering technique used.

For the proposed SSM model, Fig. 8(a) shows TER

plotted as a function of

and L parameters. We ob-

serve that the TER depends mainly on

parameter.

Higher values of

tend to merge all the keypoints, and

the lower values tend to introduce multiple keypoints

for one “star” shape. We also observe that our method

provides improved robustness to MIR, for example,

negligible MIs for the different parameter combina-

tions used, which is not the case with the radius of

cylinder used in (Utasi and Benedek, 2011; Utasi and

Benedek, 2013).

Fig. 8(b) demonstrates the Precision/Recall curves

for different L values. One of the intuitions of the

SSM is the reasonable reduction of FDR, therefore

we present the ROC curves in Fig. 8(c). It should be

noted that the number of primitives, including those

missed, vary for each value of

. Therefore, the Preci-

sion/Recall and ROC curves demonstrate the overall

performance of the shape matching, demonstrated by

ﬁxing the primitives and varying the decision bound-

aries whereas the Tab. 1 and Fig. 8(a) present the over-

all performance of our system.

4 CONCLUSION

We have presented a multi-camera system to robustly

detect people using the knowledge of the scene geom-

etry. We employ a well-known technique that projects

and merges all the views on the ground plane and

the planes parallel to it, called the synergy map. The

moving objects produce signiﬁcant values in the syn-

ergy map, and also a particular shape around it. The

popular solution to threshold the synergy map has a

drawback: ghosts are detected at locations where sev-

eral shapes induced by different people overlap. This

article proposes a solution to avoid this drawback, so

at each candidate detection, we verify if the particular

shape for an ideal person is present. This idea has

been implemented, and we focus on the two tasks in

this article: (i) how to ﬁnd the points corresponding

to potential candidates; (ii) which tolerance can be ac-

cepted when studying the shape around the candidate

detection. Quantitative results on a challenging dataset

demonstrate the performance of this approach, includ-

ing a comparison with the state-of-the-art techniques.

In the future, we propose to extend this approach by

performing 3D modelization of the shape produced by

a moving person.

ACKNOWLEDGEMENTS

The research leading to these results has received fund-

ing from the French FUI framework DeGIV.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

534

REFERENCES

Beucher, S. and Meyer, F. (1993). The morphological ap-

proach to segmentation: the watershed transformation.

Mathematical morphology in image processing. Opti-

cal Engineering, 34:433–481.

Cristani, M., Farenzena, M., Bloisi, D., and Murino,

V. (2010). Background subtraction for automated

multisensor surveillance: A comprehensive review.

EURASIP Journal on Advances in Signal Processing,

2010(1):343057.

Doll

ar, P., Wojek, C., Schiele, B., and Perona, P. (2012).

Pedestrian detection: An evaluation of the state of the

art. PAMI, 34.

Eshel, R. and Moses, Y. (2010). Tracking in a dense crowd

using multiple cameras. International Journal of Com-

puter Vision, 88(1):129–143.

Evans, M., Li, L., and Ferryman, J. (2012). Suppression of

detection ghosts in homography based pedestrian detec-

tion. In Advanced Video and Signal-Based Surveillance

(AVSS), 2012 IEEE Ninth International Conference on,

pages 31–36.

Fleuret, F., Berclaz, J., Lengagne, R., and Fua, P. (2008).

Multicamera people tracking with a probabilistic occu-

pancy map. Pattern Analysis and Machine Intelligence,

IEEE Transactions on, 30(2):267–282.

Fraley, C. and Raftery, A. E. (2007). Bayesian regulariza-

tion for normal mixture estimation and model-based

clustering. J. Classif., 24(2):155–181.

Khan, S. and Shah, M. (2009). Tracking multiple occluding

people by localizing on multiple scene planes. Pattern

Analysis and Machine Intelligence, IEEE Transactions

on, 31(3):505–519.

Li, Y., Wu, B., and Nevatia, R. (2008). Human detection by

searching in 3d space using camera and scene knowl-

edge. In Pattern Recognition, 2008. ICPR 2008. 19th

International Conference on, pages 1–5.

Liu, J., Collins, R. T., and Liu, Y. (2013). Robust auto-

calibration for a surveillance camera network. IEEE

Winter Conference on Applications of Computer Vision,

0:433–440.

Manning, C. D., Raghavan, P., and Sch

utze, H. (2008). In-

troduction to Information Retrieval. Cambridge Uni-

versity Press, New York, NY, USA.

Mehmood, M. O., Ambellouis, S., and Achard, C. (2014).

Ghost pruning for people localization in overlapping

multicamera systems. In VISAPP (2), pages 632–639.

PETS (2009). Pets dataset: Performance

evaluation of tracking and surveillance.

http://www.cvg.rdg.ac.uk/PETS2009/a.html. [Online].

Ren, J., Xu, M., and Smith, J. (2012). Pruning phantom de-

tections from multiview foreground intersection. In Im-

age Processing (ICIP), 2012 19th IEEE International

Conference on, pages 1025–1028.

Scott, D. W. (1992). Multivariate Density Estimation: The-

ory, Practice, and Visualization. Wiley, 1 edition.

Sokal, R. R. and Michener, C. D. (1958). A statistical method

for evaluating systematic relationships. University of

Kansas Scientiﬁc Bulletin, 28:1409–1438.

Tsai, R. Y. (1992). Radiometry. chapter A Versatile Camera

Calibration Technique for High-accuracy 3D Machine

Vision Metrology Using Off-the-shelf TV Cameras and

Lenses, pages 221–244. Jones and Bartlett Publishers,

Inc., USA.

Utasi, A. and Benedek, C. (2011). A 3-d marked point pro-

cess model for multi-view people detection. In Com-

puter Vision and Pattern Recognition (CVPR), 2011

IEEE Conference on, pages 3385–3392.

Utasi, A. and Benedek, C. (2013). A bayesian approach on

people localization in multicamera systems. Circuits

and Systems for Video Technology, IEEE Transactions

on, 23(1):105–115.

Yao, J. and Odobez, J. (2007). Multi-layer background

subtraction based on color and texture. In Computer

Vision and Pattern Recognition, 2007. CVPR ’07. IEEE

Conference on, pages 1–8.

LaunchTheseManhuntsShapingtheSynergyMapsforMulti-cameraDetection

535