PEOPLE DETECTION AND RE-IDENTIFICATION

FOR MULTI SURVEILLANCE CAMERAS

Etienne Corvee, Slawomir Bak and Franc¸ois Br´emond

Pulsar Team, INRIA, Sophia Antipolis, France

Keywords:

People Detection, People Tracking, People Re-identiﬁcation, Local Binary Pattern, Mean Riemannian

Covariance.

Abstract:

Re-identifying people in a network of non overlapping cameras requires people to be accurately detected and

tracked in order to build a strong visual signature of people appearances. Traditional surveillance cameras

do not provide high enough image resolution to iris recognition algorithms. State of the art face recognition

can not be easily applied to surveillance videos as people need to be facing the camera at a close range. The

different lighting environment contained in each camera scene and the strong illumination variability occurring

as people walk throughout a scene induce great variability in their appearance. In addition, people images

occlud each other onto the image plane making people detection difﬁcult to achieve. We propose a novel

simpliﬁed Local Binary Pattern features to detect people, head and faces. A Mean Riemannian Covariance

Grid (MRCG) is used to model appearance of tracked people to obtain highly discriminative human signature.

The methods are evaluated and compared with the state of the art algorithms. We have created a new dataset

from a network of 2 cameras showing the usefulness of our system to detect, track and re-identify people using

appearance and face features.

1 INTRODUCTION

Recently, the person re-identiﬁcation problem be-

came one of the most important tasks in video surveil-

lance. Only knowledge about identities of tracked

persons can allow a system to fully extract seman-

tic information about scene activities. The human re-

identiﬁcation problem can be deﬁned as a determi-

nation whether a given person of interest has already

been observed over a network of cameras.

Haar features have been studied intensely for the

detection of objects, in particular for face detection

(Viola and Jones, 2004). One another major fea-

ture used for object detection is provided by HOG

as evaluated in (Dollar et al., 2009). Pedestrians,

faces and bicycles are successfully detected when rep-

resented by HOG (Dalal and Triggs, 2005a; Adam

et al., 2006). A boosting technique is often used to

model and rapidly detect objects (Laptev, 2006) such

as humans (Zhu et al., 2006). SVM coupled with

HOG is also often used (Dalal and Triggs, 2005a) for

this task. Although Covariance features can be com-

putationally expensive to estimate, they have strong

discriminative powers. Tuzel and al. (Tuzel et al.,

2008) use a Logiboost algorithm on Riemannian man-

ifolds where Covariance features in a Riemannian

geometry are trained allowing the classiﬁcation of

pedestrians.

Many recent papers use body parts to enhance

people detection performance. There are many ways

to combine body parts; for instance Mohan et al. (Mo-

han et al., 2001) studied different voting combination

of body parts classiﬁers. In (Mikolajczyk et al., 2004),

Mikolajczyk et al. use 7 body part detectors inde-

pendently trained to better detect humans using SIFT-

like descriptors. Hussein and Porikli (Hussein et al.,

2009) introduce the notion of deformable features in

a Logiboost algorithm to allow body parts to have non

ﬁxed locations in a people image template. More re-

cently, high performances were obtained by (Huang

and Nevatia, 2010) using a highly trained set of gran-

ules.

There is a natural consequence of an inven-

tion of robust human detection algorithms to ex-

tend approaches for re-identiﬁcation purposes. The

appearance-based re-identiﬁcation techniques were

focused on associating pairs of images, each con-

taining one instance of individual. These methods

are named single-shot approaches (Bak et al., 2010b;

Park et al., 2006; Wang et al., 2007) and until now

Corvee E., Bak S. and Brémond F..

PEOPLE DETECTION AND RE-IDENTIFICATION FOR MULTI SURVEILLANCE CAMERAS.

DOI: 10.5220/0003808600820088

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2012), pages 82-88

ISBN: 978-989-8565-03-7

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

they were the most popular techniques. Currently re-

searches try to improve identiﬁcation accuracy by in-

tegrating information over many images. The group

of methods which employs multiple images of the

same person as training data is called multiple-shot

approaches.

As to single-shot approaches, in (Park et al.,

2006) the clothing colour histograms taken over the

head, shirt and pants regions together with the ap-

proximated height of the person were used as the

discriminative feature. Similarly, clothing segmen-

tation together with facial features (Gallagher and

Chen, 2008) were employed to re-identify individu-

als. Shape and appearance context model is proposed

in (Wang et al., 2007). A pedestrian image is seg-

mented into regions and their colour spatial informa-

tion is registered into a co-occurrence matrix. This

method works well if the system considers only a

frontal viewpoint. For more challenging cases, where

viewpoint invariance is necessary, the ensemble of lo-

calized features (ELF) (Gray and Tao, 2008) has been

proposed. Instead of designing a speciﬁc feature for

characterizing people appearance, a machine learning

algorithm constructs a model that provides maximum

discriminability by ﬁltering a set of simple features.

Enhancement of discriminative power of each indi-

vidual signature with respect to the others was also

the main issue in (Lin and Davis, 2008). Pairwise

dissimilarity proﬁles between individuals have been

learned and adapted into nearest neighbour classiﬁca-

tion. Similarly, in (Schwartz and Davis, 2009), a rich

set of feature descriptors based on colour, textures and

edges has been used to reduce the amount of ambigu-

ity among human class. The high-dimensional signa-

ture was transformed into a low-dimensional discrim-

inant latent space using a statistical tool called Par-

tial Least Squares (PLS) in one-against-all scheme.

Nevertheless in both methods, an extensive learning

phase based on the pedestrians to re-identify is neces-

sary to extract discriminative proﬁles what makes the

approaches non-scalable. The person re-identiﬁcation

problem has been reformulated as a ranking problem

in (Prosser et al., 2010). The authors presented exten-

sive evaluation of learning approaches and show that

a ranking relevance based model can improve the re-

liability and accuracy.

Concerning multiple-shot approaches, in (Gheis-

sari et al., 2006) the spatiotemporal graph was

generated for ten consecutive frames for grouping

spatiotemporally similar regions. Then, clustering

method is applied to capture the local descriptions

over time and improve matching accuracy. In (Bak

et al., 2010a), the AdaBoost was applied to extract the

most discriminative and invariant Haar-like features.

Here, again one-against-all learning scheme was used

to catch human dissimilarities. In (Farenzena et al.,

2010), the authors proposed to combine three fea-

tures: 1) chromatic content (HSV histogram); 2) max-

imally stable colour regions (MSCR) and 3) recurrent

highly structured patches (RHSP). The extracted fea-

tures were weighted by the distance with respect to

the vertical axis to minimize effects of pose varia-

tions. Recurrent patches were also proposed in (Baz-

zani et al., 2010). Epitome analysis was used to ex-

tract highly informative patches form the set of im-

ages.

2 OVERVIEW

Based on our previously evaluated appearance-based

people re-identiﬁcation system in (Bak et al., 2011),

we here test our system in a 2 camera network asso-

ciating people, head and face. The ﬁrst task is to de-

tect people using simpliﬁed LBP features in section

3. Tracked objects are extracted from a video using a

temporal trajectory analysis algorithm (Avanzi et al.,

2001) which are fed to the appearance based people

re-identiﬁcation system in section 4. We propose a

simple face recognition algorithm in section 5. Re-

identiﬁcation results are given in section 6. A new

database was created to simulate a 2 cameras network

with non overlapping views. This database is referred

to as the TSP database and contains 23 persons walk-

ing back and forth in both cameras as illustrated in

ﬁgure 1. Two tracked persons with heads and faces

are shown in camera 1 and 2 of the TSP camera setup

in ﬁgure 2.

Figure 1: TSP camera setup: 2 non overlapping cameras in

2 joint corridors.

PEOPLE DETECTION AND RE-IDENTIFICATION FOR MULTI SURVEILLANCE CAMERAS

Figure 2: Tracked people, head and faces in up: TSP cam-

era 1 and down: TSP camera 2.

3 PEOPLE DETECTION USING

SIMPLIFIED LOCAL BINARY

PATTERN

The standard LBP operator extract feature vector of

size 256 (Trefny and Matas, 2010). We have imple-

mented a new simpliﬁed version of the LBP operator

by reducing its dimensionality to 16. which we will

refer to as the SLBP for the rest of the paper. This

SLBP is calculated from 4 cells as illustrated in im-

age 3. A pixel intensity average v

c,w,h

is calculated

for each of the 4 cells c, c = [1 : 4] as deﬁned in equa-

tion 1 where E represents the expectation (average)

function over pixel locations (x, y).

c,w,h

= E

′



I(x

′

, y

′

)



(1)

c =











1 : x

′

= [x : x + h], y

′

= [y : y + w]

2 : x

′

= [x : x + h], y

′

= [y + w : y+ 2w]

3 : x

′

= [x + h : x+ 2h], y

′

= [y : y + w]

4 : x

′

= [x + h : x+ 2h], y

′

= [y + w : y+ 2w]

The SLBP feature is then calculated from 4 mean

pixel intensity differences ∆v

as follows:

SLBP

w,h

(x) =

∑

q=0

s(∆v

, s(.) =



1 if . > V

0 else

∆v

= v

− v

(q, i, j) = {(0, 1, 0), (1, 3, 2), (2, 2, 0), (3, 3, 1)}

An Adaboost training scheme is adopted to train cell

features across the a people image database. The

training is performed on a multiple scale approach by

varying the SLBP cell dimensions:

(w, h) = { (1, 1)(1, 2)..(1, 8), (2, 1)..(8, 8)}

We apply the same training algorithm on head and

face image datasets for head and face detection. The

training for people detection is done using 7K positive

images (5K from NICTA, 1K from MIT and 1K from

INRIA people training dataset) and 50 background

negative PAL images. Head were trained with 1K

positive images (cropped head images from INRIA

and TUD people training dataset) and 10 background

negative images. Faces were trained with the standard

CMU face training database.

The traditional Adaboost technique is used to train

SLBP features. We reach a frame rate of 1fps for

a minimum size of 160 pixels for people height in

PAL images and when the scan is performed with an

increase rate of 10% of its actual scanning window

width and height. Simple rules are applied to the de-

tected candidates in order to fuse overlapping candi-

dates and eliminate stand alone noisy candidates: ob-

jects are merged if they overlap each other with a min-

imum union-over-intersectionratio of 50% and a ﬁnal

object requires a minimum number of 2 overlapping

ones.

Figure 3: Simpliﬁed LBP operator.

4 HUMAN RE-IDENTIFICATION

Recently, the person re-identiﬁcation problem be-

came one of the most important tasks in video surveil-

lance. There is a natural consequence of an in-

vention of robust human detection algorithms to ex-

tend approaches for recognition purposes. Person re-

identiﬁcation can be considered on different levels

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

Figure 4: Covariances gathered from tracking results are

used to compute the MRC using Riemannian manifold

space (depicted with the surface of the sphere).

depending on information cues which are currently

available in the system. Biometrics such as face, iris

or gait can be used to recognize identities. Never-

theless, in most video surveillance scenarios such de-

tailed information is not available due to video low-

resolution or difﬁcult segmentation (crowded envi-

ronments, e.g. airports, metro stations). Therefore

a robust modelling of a global appearance of an indi-

vidual is necessary to re-identify a given person of

interest. In these identiﬁcation techniques (named

appearance-based approaches) clothing is the most

reliable information about an identity of an individ-

ual (there is an assumption that individuals wear the

same clothes between different sightings). The model

of an appearance has to handle differences in illumi-

nation, pose and camera parameters to allow match-

ing appearances of the same individual observed in

different cameras.

In (Bak et al., 2011), a highly discriminative hu-

man signature, called Mean Riemannian Covariance

Grid (MRCG), is proposed. This human signature

has been designed to deal with low resolutions im-

ages and crowded environments where more special-

ized techniques (e.g. based on body parts detectors)

might fail. Here, dense descriptors philosophy (Dalal

and Triggs, 2005b) is combined with extremely effec-

tiveness of the covariance descriptor. First, an image

is divided into a dense grid structure with overlapping

spatial square regions (cells). Such dense representa-

tion makes the signature robust to partial occlusions

and it contains a relevant information about spatial

correlations between cells. The authors (Bak et al.,

2011) take advantage of the tracking and detection re-

sults combining information from multiple images.

Let C

, . . . ,C

be a set of covariance matrices ex-

tracted during tracking of N frames corresponding to

image square regions at position of the cell p. The

MRC is deﬁned as the mean covariance of these co-

variance matrices computed using Riemannian space

(see Fig. 4). The mean covariance matrix as an intrin-

sic average blends appearance information from mul-

tiple images. This mean covariance matrix keeps not

only information about features distribution but also

carries out essential cues about temporal changes of

the appearance related to the position of the cell p.

All MRC cells compose a full grid, named as Mean

Riemannian Covariance Grid (MRCG).

In our surveillance system, this appearance-based

descriptor is used to match the same appearances be-

tween different camera views.

5 FACE RECOGNITION

5.1 Face Visual Signature Extraction

Similarly to (Huang and Nevatia, 2010) where people

body parts are manually located before being inde-

pendently trained, we manually choose 4 facial parts

(p = [1 : 4]) to model a face i visual signature S

(i)

which are the left eye, right eye, nose and mouth:

(i)

; p = [1 : 4]

(2)

Each signature H

(i)

is represented by a set of SLBP

histograms h

(i)

p,w,h

( f):

(i)

p,w,h

( f) ; (w, h) = {(1, 1)..(1, 8)..(8, 8)}

(i)

p,w,h

( f) =

∑

x∈P

′

(SLBP

(i)

w,h

(x), f) (3)

′

(a, b) =



1 if a = b

0 else

(4)

where P

represent the set of pixels within the pre-

deﬁned facial part area. Each histogram h is nor-

malised over the SLBP feature value f = [1 : 16].

5.2 Matching Faces

Two faces i and j are matched if their signatures S

(i)

and S

( j)

are similar enough or in other terms if a dis-

tance measure between the two signatures is below

a threshold. This similarity distance D(i, j) is cal-

culated by the mean similarity distance ∆H

(i, j) of

the 4 facial part signatures H

deﬁned in equation 2.

Two facial parts similarity is calculated by a weighted

mean of SLBP histogram differences as follows.

D(i, j) = E

(∆H

(i, j)) (5)

∆H

(i, j) =

∑

(w,h)

w,h





(i)

p,w,h

( f) − h

( j)

p,w,h

( f)





∑

w,h

where cell weight α is inversely proportionally to the

cell lateral distance.

PEOPLE DETECTION AND RE-IDENTIFICATION FOR MULTI SURVEILLANCE CAMERAS

6 EVALUATION

6.1 People Detection

We have evaluated our people detection algorithm on

the test human dataset provided by INRIA against

state of the art algorithms which we refer as HOG

(Dalal and Triggs, 2005b) and LBP-HOG (Wang

et al., 2009). The INRIA human dataset is composed

of 1132 human images and 453 images of background

scenes containing no humans. The results are dis-

played in ﬁgure 5 which shows that we obtain slightly

better performances than the HOG-LBP technique in

terms of missed detection rate vs. FPPI i.e. False Pos-

itive Per Image. In this ﬁgure, two extreme function-

ing modes could be chosen: approximately 2 noisy

detections are obtained every 1000 background im-

ages for 50% true positive detections or 1 noisy de-

tection every 2 frames for a detection rate of approxi-

mately 88%.

Figure 5: People detection evaluation: false positive per im-

age vs. miss detection rate for the INRIA test database.

6.2 Appearance based People

Re-identiﬁcation

Using the TSP dataset described in section 2, three

kind of people are noted:

• type A - people wearing the same clothes in

recording sessions.

• type B - people slightly change their appearances

from one recording session to the other. For exam-

ple, someone unzipping his/her jacket or someone

taking his/her scarf off as illustrated by the top im-

ages of ﬁgure 6.

• type C - people with great change of appearance

such as a person adding a coat with a hat as illus-

trated by the bottom images of ﬁgure 6.

We obtain 50 possible comparisonof appearancesthat

we evaluated in section 6.2.1 and 6.2.2. We can note

that people detection and tracking results are noisy

which makes the people re-identiﬁcation task more

challenging.

Figure 6: Top row: type B - people appearances are suc-

cessfully re-identiﬁed despite their weak change of cloth-

ing. Bottom row: type C - people appearances are too dis-

tant to be re-identiﬁed.

6.2.1 Same Camera Re-identiﬁcation Scenario

In this scenario, we aim to re-identify people in a

camera who were previously seen in this same cam-

era. The tracked people appearance signatures are

recorded in a database during the ﬁrst session. During

the second session, the appearance signatures of all

the tracked people are compared with the database.

We have plotted in ﬁgure 7 the appearance match-

ing distance between the same person appearances

for type A, B and C people (deﬁned above). The

results show that people type A and B are success-

fully retrieve with rank 1, except for 1 person with

rank 2, when matching distances are below 2.1. Type

C people (i.e. strong change of appearance) are not

re-identify. However, a uncertainty zone exists for a

matching distance between 1.75 and 2.1.

Figure 7: Evaluation of appearance matching distance in the

same camera scenario.

VISAPP 2012 - International Conference on Computer Vision Theory and Applications

6.2.2 Different Camera Re-identiﬁcation

Scenario

Figure 8: CMC curves obtained on TSP datasets.

In this section the evaluation of re-identiﬁcation ap-

proach is presented in a 2 camera network. The

performance is shown using the Cumulative Match-

ing Characteristic (CMC) curve suggested in (Gray

et al., 2007) as the validation method for the re-

identiﬁcation problem. The CMC curve represents

the expectation of ﬁnding the correct match in the top

matches. We evaluate the re-identiﬁcation in the fol-

lowing way:

• Mono appearance mode - Frontal and back view

of people tracked in camera 1 have their appear-

ance signatures extracted separately.

• Multi appearance mode - Both frontal and back

view of people tracked in camera 2 have their ap-

pearance signatures extracted.

Every signature from camera 1 is used as a query

to the database (multiple-appearance signatures ex-

tracted from camera). The CMC curves for back and

frontal view signatures are presented in Fig. 8. The

results show that despite automatic inaccurate people

detection and the different lighting conditions of the 2

cameras, the system shows promising people appear-

ance based re-identiﬁcation results.

6.3 Face Recognition

The Olivetti Research Lab i.e. ORL database is

used to evaluate our simple face recognition algo-

rithm using SLBP features. 40 persons constitute this

database with 10 pictures for each subject. All face

pictures are taken by the same camera. Most state of

the art algorithm compare face recognition rate using

one picture at a time. In our scenario, we obtain tracks

of faces for each person which we aim to compare. To

simulate our 2 cameras system, we split the database

Table 1: Face recognition rates.

technique recognition rate %

ORL TSP1 TSP2

DAISY 98.2

PDSIFT 95.5

SIFT Grid 95.2

SLBP 85 91 100

SLBP(rank 1) 98 100 100

SLBP(rank 3) 100 100 100

in two: 5 images are used as our face database col-

lected from one camera and the remaining 5 images

are used as a request from camera 2. The images are

challenging as they contain faces of different kind of

poses and facial expressions.

Table 1 shows the recognition rates comparison

with 3 different techniques. Using an empirically

set threshold (of 0.001), our system recognises only

85% of faces correctly whereas the DAISY technique

(Velardo and Dugelay, 2010) recognises 98% of the

faces correctly. Although our system is far from be-

ing the most performing system, it recognises 98% of

the people when using a rank score of 1 and 100% for

a rank of 3. In other terms, there is 98% chance that

the face giving the closest distance D (in equation 5)

is the correct face.

Using the TSP dataset, we obtained similar per-

formances. Among the 23 persons present in the TSP

database, only 11 people faced long enough camera

1 for a valid trajectory to be obtained: face images

of less than 50x50 pixels were not taken into account

(due to motion noise and image quality artefacts) and

face tracks containing less than 5 face images were

discarded.

In table 1, TSP1 and TSP2 refers to the same cam-

era (camera 1 and camera 2 respectively) recognition

scenario where half of a person face images is saved

in the database and the other half as the request in-

formation. The results show that this simple SLPB

face recognition technique is useful enough for such

a database.

7 CONCLUSIONS

We have presented a multi-cue people re-

identiﬁcation framework. A novel simpliﬁed

LBP feature is proposed to detect people, head and

faces using the Adaboost training scheme. We obtain

state of the art performance for people detection.

We have also extended our approach on appearance

based matching method to multi appearance based

PEOPLE DETECTION AND RE-IDENTIFICATION FOR MULTI SURVEILLANCE CAMERAS

people re-identiﬁcation. The proposed system tracks

people and their faces allowing the user to associate

an appearance with a face. In most networks,

cameras cannot provide the full people appearance

view (e.g. strong occlusion) and faces are often not

visible or only partly visible. Our proposed approach

would allow a user to scan throughout a network of

surveillance cameras the best matching candidates

and to be able to track people of interest throughout

this network.

ACKNOWLEDGEMENTS

We would like to thank the ANR project ’VideoId’

who partially founded this work and the following

partners of the project: Biometrics Groups at TELE-

COM SudParis, Multimedia Image processing Group

of Eurecom and T3S (Thales Security Systems and

Solutions S.A.S.).

REFERENCES

Adam, A., Rivlin, E., and Shimshoni, I. (2006). Robust

fragment-based tracking using integral histogram. In

Computer Vision and Pattern Recognition - CVPR.

Avanzi, A., Bremond, F., and Thonnat, M. (2001). Track-

ing multiple individuals for video communication. In

In IEEE Proc. of International Conference on Image

Processing, Thessaloniki (Greece).

Bak, S., Corvee, E., Bremond, F., and Thonnat, M. (2010a).

Person re-identiﬁcation using haar-based and dcd-

based signature. In 2nd Workshop on AMMCSS.

Bak, S., Corvee, E., Bremond, F., and Thonnat, M. (2010b).

Person re-identiﬁcation using spatial covariance re-

gions of human body parts. In AVSS.

Bak, S., Corvee, E., Bremond, F., and Thonnat, M. (2011).

Multiple-shot human re-identiﬁcation by mean rie-

mannian covariance grid. In AVSS.

Bazzani, L., Cristani, M., Perina, A., Farenzena, M.,

and Murino, V. (2010). Multiple-shot person re-

identiﬁcation by hpe signature. In ICPR, pages 1413–

1416.

Dalal, N. and Triggs, B. (2005a). Histograms of Oriented

Gradients for Human Detection. In Computer Vision

and Pattern Recognition - CVPR.

Dalal, N. and Triggs, B. (2005b). Histograms of oriented

gradients for human detection. In CVPR, pages 886–

893.

Dollar, P., Wojek, C., Schiele, B., and Perona, P. (2009).

Pedestrian detection: a benchmark. In CVPR.

Farenzena, M., Bazzani, L., Perina, A., Murino, V., and

Cristani, M. (2010). Person re-identiﬁcation by

symmetry-driven accumulation of local features. In

CVPR, pages 2360–2367.

Gallagher, A. C. and Chen, T. (2008). Clothing cosegmen-

tation for recognizing people. In CVPR, pages 1–8.

Gheissari, N., Sebastian, T. B., and Hartley, R. (2006). Per-

son reidentiﬁcation using spatiotemporal appearance.

In CVPR, pages 1528–1535.

Gray, D., Brennan, S., and Tao, H. (2007). Evaluating Ap-

pearance Models for Recognition, Reacquisition, and

Tracking. PETS.

Gray, D. and Tao, H. (2008). Viewpoint invariant pedestrian

recognition with an ensemble of localized features. In

ECCV, pages 262–275.

Huang, C. and Nevatia, R. (2010). High performance object

detection by collaborative learning of joint ranking of

granules features. IEEE - CVPR 2010.

Hussein, M., Porikli, F., and Davis, L. (2009). Object de-

tection via boosted deformable features. IEEE Inter-

national Conference on Image Processing (ICIP).

Laptev, I. (2006). Improvements of object detection using

boosted histograms. In Proceedings of the British Ma-

chine Vision Conference.

Lin, Z. and Davis, L. S. (2008). Learning pairwise dis-

similarity proﬁles for appearance recognition in visual

surveillance. In ISVC, pages 23–34.

Mikolajczyk, K., Schmid, C., and Zisserman, A. (2004).

Human detection based on a probabilistic assembly of

robust part detectors. In ECCV.

Mohan, A., Papageorgiou, C., and Poggio, T. (2001).

Example-based object detection in images by com-

ponents. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 23:349–361.

Park, U., Jain, A., Kitahara, I., Kogure, K., and Hagita, N.

(2006). Vise: Visual search engine using multiple net-

worked cameras. In ICPR, pages 1204–1207.

Prosser, B., Zheng, W.-S., Gong, S., and Xiang, T. (2010).

Person re-identiﬁcation by support vector ranking. In

BMVC, pages 21.1–21.11.

Schwartz, W. R. and Davis, L. S. (2009). Learning discrim-

inative appearance-based models using partial least

squares. In SIBGRAPI, pages 322–329.

Trefny, J. and Matas, J. (2010). Extended Set of Local Bi-

nary Patterns for Rapid Object Detection. In Com-

puter Vision Winter Workshop 2010 - CVWW10.

Tuzel, O., Porikli, F., and Meer, P. (2008). Pedestrian detec-

tion via classiﬁcation on riemannian manifolds. PAMI,

30(10).

Velardo, C. and Dugelay, J. (2010). Face recognition

with daisy descriptors. In MM’10 and Sec’10, ACM

SIGMM Multimedia and Security Workshop, Septem-

ber 9-10, Rome, Italy.

Viola, P. and Jones, M. (2004). Robust real-time face detec-

tion. In International Journal of Computer Vision.

Wang, X., Doretto, G., Sebastian, T., Rittscher, J., and Tu,

P. (2007). Shape and appearance context modeling. In

ICCV, pages 1–8.

Wang, X., Han, T., and Yan, S. (2009). An hog-lbp human

detector with partial occlusion handling. ICCV09 -

International Conference on Computer Vision.

Zhu, Q., Avidan, S., Yeh, M., and Cheng, K. (2006). Fast

human detection using a cascade of histograms of

oriented gradients. In Computer Vision and Pattern

Recognition - CVPR.

VISAPP 2012 - International Conference on Computer Vision Theory and Applications