HDR Imaging for Enchancing People Detection and Tracking in

Indoor Environments

Panagiotis Agraﬁotis, Elisavet K. Stathopoulou, Andreas Georgopoulos and Anastasios Doulamis

Laboratory of Photogrammetry, School of Rural and Surveying Engineering, National Technical University of Athens,

15780 Zografou Athens, Greece

Keywords:

HDR Imaging, People Detection, People Tracking.

Abstract:

Videos and image sequences of indoor environments with challenging illumination conditions often capture

either brightly lit or dark scenes where every single exposure may contain overexposed and/or underexposed

regions. High Dynamic Range (HDR) images contain information that standard dynamic range ones, often

mentioned also as low dynamic range images (SDR/LDR) cannot capture. This paper investigates the contri-

bution of HDR imaging in people detection and tracking systems. In order to evaluate this contribution of the

HDR imaging in the accuracy and robustness of pedestrian detection and tracking in challenging indoor visual

conditions, two state of the art trackers of different complexity were implemented. To this direction data were

collected taking into account the requirements and real-life indoor scenarios and HDR frames were produced.

The algorithms were applied to the SDR data and their corresponding HDR data and were compared and eval-

uated for their robustness and accuracy in terms of precision and recall. Results show that that the use of HDR

images enhances the performance of the detection and tracking scheme, making it robust and more reliable.

1 INTRODUCTION

Videos and image sequences of indoor environments

with challenging illumination conditions often cap-

ture brightly lit or dark scenes where every single

exposure may contain overexposed and/or underex-

posed regions. This common phenomenon is usu-

ally the result of human constructions (buildings, win-

dows etc.) and trees that inﬂuence negatively the dig-

ital representation of the scene. Essential information

loss, noise and saturation degrade image processing

and computer vision algorithms that rely on accurate

feature detection.

High Dynamic Range (HDR) images contain in-

formation that standard dynamic range ones, often

mentioned in the bibliography also as low dynamic

range images (SDR/LDR) cannot capture. A standard

sensor is designed to depict a limited range of inten-

sities at the same frame, in contrast with the human

eye adaptation system. Thus, in extreme illumination

conditions, overexposed and/or underexposed pixels

may occur, causing serious information loss. HDR

images can be captured either by specially designed

sensors with extended dynamic range or by merging

multiple SDR images of the same scene with different

exposure time settings (Reinhard et al., 2005).

This study investigates the contribution of HDR

imaging in people detection and tracking systems

through the implementation of two detection and

tracking systems of different complexity. The paper is

organised as follows: Section 2 is a review of related

previous work concerning HDR imaging and feature

detection and tracking algorithms. Section 3 analyses

our approach, followed by experimental results with

captured datasets in Section 4 and conclusions in Sec-

tion 5.

2 RELATED WORK

The current state-of-the-art in such computer vision

algorithms study constrained visual environments like

the ones within a research laboratory or of real-like

sequences of one or few objects. To handle these

bottlenecks, in the recent years, several research ef-

forts have been also published using complicated vi-

sual environments, like the ones of outdoors surveil-

lance of multiple persons (pedestrians), crowded con-

ditions, aerial monitoring, teleconferences rooms,

sports events, and daily activities within houses.

Creating an HDR image by merging multiple

views of the same scene captured with different ex-

posure values is commonly used in the bibliography.

623

Agraﬁotis P., Stathopoulou E., Georgopoulos A. and Doulamis A..

HDR Imaging for Enchancing People Detection and Tracking in Indoor Environments.

DOI: 10.5220/0005456706230630

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (MMS-ER3D-2015), pages 623-630

ISBN: 978-989-758-090-1

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

(Debevec and Malik, 1997; Robertson et al., 1999;

Rovid et al., 2007). These algorithms use the input

images to create the HDR image by computing the

camera response function (CRF) given the exposure

value.

HDR images are basically radiance maps whose

pixels reach a maximum of 32-bit ﬂoating point im-

age representation. Common display media can visu-

alise SDR/LDR images in a 16- or 8-bit format but

are unable to render properly images in 32-bit format.

Various compression operators to convert HDR im-

ages to displayable (LDR) ones without serious detail

loss have been developed (Reinhard et al., 2002; Du-

rand and Dorsey, 2002; Drago et al., 2003; Mantiuk

et al., 2006; Fattal et al., 2007). This process is known

as tone-mapping technique. Tone mapping operators

can be classiﬁed into global and local with respect to

their pixel region of impact. Global techniques apply

a mapping function to the entire set of pixels, while

local ones take into account the characteristics in the

neighbourhood of each pixel.

Feature detection and tracking in images and

videos is a common problem in computer vision and

it has been extensively discussed in the literature dur-

ing the previews years. First step of these approaches

is the extraction of appropriate visual descriptors.

The most commonly used detectors and descriptors

in vision based applications are SIFT, SURF, GFTT,

FAST, ORB and HOG. Scale Invariant Feature Trans-

form (SIFT) (Lowe, 2004) extracts features invariant

to image scale, rotation and translation and partially

invariant to illumination changes. Speeded Up Ro-

bust Features (SURF) (Bay et al., 2008) and Good

Features To Track (GFTT) (Shi and Tomasi, 1994),

feature detectors are based on the assessment of the

Hessian matrix. SURF also combines both detec-

tion and description but it outperforms SIFT in terms

of speed. GFTT is an improved version of Harris

corner detector (Harris and Stephens, 1988). Fea-

tures from Accelerated Segment Test (FAST) (Ros-

ten and Drummond, 2006) is based on the intensity

difference of the candidate centre pixel and its sur-

rounding neighbours lying on a circular ring about the

centre. It is beneﬁting from low computation. The

Oriented FAST and Rotated BRIEF (ORB) (Rublee

et al., 2011). ORB is a combination of FAST and

the recently developed BRIEF descriptor (Calonder

et al., 2010), being rotation invariant and resistant to

noise. Finally, HOG (Histograms of Oriented Gra-

dients) were initially described by Dalal and Triggs

(Dalal and Triggs, 2005) in the context of person de-

tection in images and videos. These descriptors are

fed as input to non-linear classiﬁers to detect the ob-

jects and track their position within the space (Liu

et al., 2011; Kaaniche and Bremond, 2009; Alahi

et al., 2010; Miao et al., 2011).

In addition, great effort has been dedicated to han-

dle object tracking as a classiﬁcation problem (Avi-

dan, 2004; Lepetit et al., 2005). Some of the ﬁrst ap-

proaches towards an adaptable classiﬁcation for ob-

ject tracking have been presented in (Collins et al.,

2005; Doulamis et al., 2003). These studies, how-

ever, do not face efﬁciently the general trade-off be-

tween model stability and adaptability. Matthews et

al. (Matthews et al., 2004), propose an updated tem-

plate algorithm that avoids the ”drifting” inherent.

Other methods exploit the semi-supervised paradigm

(Stalder et al., 2009; Grabner et al., 2008), a co-

training strategy (Tang et al., 2007), a combination

of generative and discriminative trackers (Yu et al.,

2008), or ﬁnally coupled layered visual models (Ce-

hovin et al., 2011).

The fact that local information sometimes is not

enough to make a correct decision led to the devel-

opment of global optimization trackers. Examples

are the use of a minimum cost graph matching that

runs the Hungarian algorithm (Cehovin et al., 2011;

Huang et al., 2008). Other works handle the problem

as a minimum ﬂow cost problem (Zhang et al., 2008;

Henriques et al., 2011). However, in case that the

background/foreground signiﬁcantly changes, there is

no way to estimate matches for longer time periods

forcing the algorithm to fail. The aforementioned ap-

proaches exploit Standard Dynamic Range (SDR) im-

ages or videos that limit their performance.

Although HDR imaging is a broadly used tech-

nique, a few surveys regarding feature tracking on

such images have been made. Cui et al. proved that

the performance and robustness of SIFT operator on

HDR images is superior than using SDR ones (Cui

et al., 2011). The work of Ladas et al. (Ladas et al.,

) takes advantage of HDR imaging and inverse illu-

mination based on Precomputed Radiance Transfer

(PRT) (Sloan et al., 2002). Chermak and Aouf investi-

gated HDR imaging on a larger number of state of the

art operators, SIFT, SURF, Harris, Tomasi and FAST,

resulting in indeed higher performance than SDR im-

ages (Chermak and Aouf, 2012). A subsequent study

(Chermak et al., 2014) elaborates further on this ﬁeld

by testing larger image sequences for feature tracking.

The main contribution of this paper is the use of

the HDR imagery instead of the SDR already used, in

order to eventually improve the performance and the

reliability of pedestrian detection and tracking algo-

rithms.

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

624

3 APPROACH

3.1 Tested Trackers

In order to evaluate the contribution of the High Dy-

namic Range imaging in the accuracy and robustness

of pedestrian detection and tracking in challenging in-

door visual conditions, two state of the art trackers of

different complexity were implemented.

3.2 The Complex Tracker

We adopt and modify the tracking scheme presented

in (Kokkinos et al., 2013). The algorithm operates in

real time, or at least just in time and the computer

vision tools have been developed using Intels Inte-

grated Performance Primitives tools, exploiting pro-

cessor hardware capabilities.

This exploited and modiﬁed version of the afore-

mentioned human tracker, integrates on the one hand

adaptive background models able to capture slight

modiﬁcations of the background with, on the other

hand, algorithms based on motion characteristics in

order to deﬁne accurately and precisely the areas of

the image to be considered as foreground or back-

ground.

In particular, for background modelling, local ge-

ometrically enriched mixture models were introduced

like the Gaussian Mixture Models (GMMs) in order

to deal with the problem of tracking over long peri-

ods. These models are exploiting the geometric prop-

erties of locally connected regions. Since, such an

approach is very sensitive to noise and the alteration

of the pixel values due to illumination changes, re-

ducing the background modelling performance, the

dimensionality of the hyper-space is reduced using

the concept of saliency maps. This approach adopts

a graph-based saliency map methodology so that the

most important pixels are selected (pixels in which

humans feature more than others), and these to be

fed as GMM input. This inclusion of local connec-

tivity in modelling the background content increases

robustness and tolerance to noise owing either to cam-

era defects and illumination ﬂuctuation. In order to

make the algorithm applicable in real-time applica-

tions, the background on the on-line captured video

frames is re-modelled only in high conﬁdent areas of

background. These areas are estimated using an iter-

ative motion detection scheme constrained by shape

and time properties. In particular, the foreground

object is currently detected as a moving object pre-

senting human shape constraints while retaining its

continuity in time (temporal coherency). The mo-

tion ﬁeld is detected through an iterative implementa-

tion of the Lucas-Kanade Optical Flow method (Lu-

cas et al., 1981) while a constrained shape and time

mechanism is developed in order to overcome intro-

duced noise. For accelerating the optical ﬂow algo-

rithm, is adopted an iterative implementation of the

algorithm starting with the creation of a stack of dif-

ferent image resolutions (Doulamis, 2010). This is

achieved by low-pass ﬁltering of the image content.

However, estimating the motion activity for all

pixel of a frame is a time-consuming process while

increases the computational complexity of the algo-

rithm, threatening a real-time implementation. More-

over, the application of the aforementioned technique

for all pixels would provide a large number of erro-

neous motion vectors mainly due to the high dynam-

ics of the background content and the complexity of

the visual environment. It is also clear that erroneous

estimation of motion vectors results in an inaccurate

detection, increasing false positives/negatives.

To address this difﬁculty, the implemented scheme

estimates the motion vectors on particularly selected

points on the image plane. This is done by detect-

ing good image pixels. In this approach, is imple-

mented the Good Features to Track method of Shi and

Tomasi for pixel representation as feature vectors (Shi

and Tomasi, 1994). The aforementioned scheme has

been also used for activity and behavior recognition

(Kosmopoulos et al., 2012; Voulodimos et al., 2014;

Voulodimos et al., 2012).

3.3 The HOG-based Tracker

In parallel with the already described implementation,

we test a simpler tracker based on Histograms of Ori-

ented Gradients (HOG) (Dalal and Triggs, 2005) and

a linear Support Vector Machine (SVM) classiﬁer. A

detection and tracking system based on this method-

ology is sensitive in luminosity conditions and this

makes it ideal for testing a possible HDR enchant-

ment for pedestrian detection and tracking.

Subsequently, a detailed review of Dalal and

Triggs method is given. The HOG human detector is

based on evaluating well-normalized local histograms

of image gradient orientations in a dense grid. The

main concept of this methodology is that local ob-

ject appearance and shape can often be characterized

rather well by the distribution of local intensity gradi-

ents or edge directions, even without precise knowl-

edge of the corresponding gradient or edge positions

(Dalal and Triggs, 2005). In order to construct this

dense grid of oriented gradient histograms, several

stages are implemented. Initially, gamma and colour

normalization steps are carried out to reduce the in-

ﬂuence of shadows and illumination. The intensity

HDRImagingforEnchancingPeopleDetectionandTrackinginIndoorEnvironments

625

gradient at a pixel is then computed for each colour

channel. The next stage is gradient computation at

each spatial region or (cell). A range of adjacent

cells are grouped into larger spatial blocks so as to

normalize edge contrast and illumination. The ﬁnal

step is to concatenate all of the normalized block vec-

tors over each detection window to form a ﬁnal win-

dow descriptor. HOG descriptors are computed over

a number of scales on overlapping windows in the im-

age. The high dimension of the HOG descriptors al-

lows a simple linear SVM classiﬁer to successfully

classify detection windows into non-pedestrian and

pedestrian. In order to train the a linear Support Vec-

tor Machine (SVM) classiﬁer, Dalal and Triggs select

1239 of the images as positive training examples, to-

gether with their left-right reﬂections (2478 images

in total) (Dalal and Triggs, 2005). This database in-

cludes pedestrian in many different standing/walking

poses in complex scenes. In addition, these images

include indoor and outdoor scenes taken at different

times of the day. Nevertheless, the use of HOG de-

scriptors is a memory consuming process because of

the high dimension of the descriptors. In addition, the

HOG human detector requires the pedestrians to be in

upright positions. For poses other than standing and

walking, the performance of the HOG human detector

deteriorates (Jiang et al., 2010).

4 EXPERIMENTAL RESULTS

4.1 Test Dataset

Since there are not any standard datasets including

HDR sequences, in order to evaluate the perfor-

mance of HDR imaging on pedestrians tracking

and detection algorithms, data were collected taking

into account the requirements and real-life indoor

scenarios. More speciﬁcally an indoor sequence

of a moving person acquired having challenging

illumination conditions. In more detail, a person

is walking away from the camera, moving ﬁrstly

almost in parallel with the optical axis and then

perpendicular to it. In addition, overexposed and

underexposed areas appear in the frames. The

sequences were recorded with recording rate 5 fps

and 640 × 430 resolution. Furthermore, other data

including pedestrians videos were used in order to

assess the detection and tracking performance of

the applied algorithms (Agraﬁotis et al., 2014b;

Agraﬁotis et al., 2014a).

(a) (b) (c)

Figure 1: Three consecutive frames of ±0 exposure level.

4.2 HDR Image Creation

In order to create each HDR image, 7 im-

ages were captured with different exposure level

(−3, −2, −1, 0, +1, +2, +3). Having collected the

necessary data, HDR images (radiance maps) were

created for each frame implying the algorithm pre-

sented in (Debevec and Malik, 1997), which uses the

reciprocity law in order to recover the response curve.

Since these HDR images are not displayable, Durand

and Dorsey (Durand and Dorsey, 2002) operator is

used afterwards for the tonemapping process which

calculates the corresponding LDR of the HDR image

with minimum information loss.

(a) (b)

(e) (f)

(g) (h)

Figure 2: (a)-(f) The 7 simultaneous images captured

with different exposure level (−3, −2, −1, 0, +1, +2, +3)

respectively and (g) the corresponding HDR image.

This operator decomposes the HDR image into

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

626

two layers, a base and a detail one, using edge-

preserving bilateral ﬁlter. To avoid any confusion,

from now on we will refer to the original captured

data as SDR images, while LDR images will mean

the one produced after the tonemapping procedure.

4.3 Tracking Results Comparison and

Evaluation

(a) (b)

(e) (f)

(g) (h)

(i) (j)

(k) (l)

Figure 3: Tracking results from original SDR images of ±0

level of exposure (left) and from LDR images (right) for

HOG based algorithm.

Aiming to evaluate the contribution of the HDR Imag-

ing in detection and tracking systems, initially the al-

gorithms described in Section 3.2 are applied in the

original SDR test data presented in Section 4.1 and

their corresponding LDR images of HDR data as to

be compared and evaluated for their robustness and

accuracy in terms of precision and recall.

Figures 3 and 4 present examples of detection and

tracking results of the implemented algorithms using

original SDR images of ±0 level of exposure on the

left column and using LDR images on the right col-

umn. It is clear that the use of HDR images (in fact

their corresponding LDR ones) enhances the perfor-

mance of the detection and tracking scheme making it

robust and more reliable. In more detail, HDR imag-

ing in dark indoor visual environments with overex-

posed and underexposed images areas, improves the

accuracy of the tracker by reducing greatly false pos-

itives and increasing true positives. As it is observed,

in original SDR images are detected in average 90%

more false people detections than in LDR images.

(a) (b)

(e) (f)

(g) (h)

(i) (j)

(k) (l)

Figure 4: Tracking results from original images of ±0 level

of exposure (left) and from LDR images (right) for the com-

plex algorithm (Kokkinos et al., 2013).

The graph in Figure 5 presents the average preci-

HDRImagingforEnchancingPeopleDetectionandTrackinginIndoorEnvironments

627

sion and recall from all tested data and both imple-

mented detection and tracking schemes. In terms of

precision, the SDR dataset achieves 30.3% and the

HDR dataset 75%. It seems that the use of HDR im-

ages leads in more than doubling the percentage of

precision term. Such increase in precision means that

the tested algorithms returned substantially more rel-

evant results than irrelevant through the use of HDR

data.

Figure 5: Average precision and recall from all tested data

and both implemented detection and tracking schemes.

In terms of recall, the SDR dataset achieves 50%

and the HDR dataset 71%. These results denote

that the algorithms returned most of the relevant re-

sults. As before, signiﬁcant improvement of the per-

formance of the algorithms after using HDR imagery

is observed here.

5 CONCLUSIONS

We have proven and presented that HDR imaging

could enhance the performance and the reliability of

pedestrian detection and tracking algorithms. Our ex-

perimental results show that the HDR imagery in-

creases over 100% the precision term and about 50%

the recall term. This technique can reduce the use of

IR cameras and consequently reduce the cost of such

systems.

Our future work includes extending our experi-

ments to implement more state of the art trackers

of different complexity combining them with more

HDR creation algorithms. In addition, our future

work includes capturing datasets from outdoor envi-

ronments with reduced illumination, foggy or snowy

areas where the latter are characterized for containing

overexposed and underexposed regions. Finally, se-

quences from urban environments are to be tested in

order to deal with the problem of building shadows

and their underexposed areas.

ACKNOWLEDGEMENTS

The research leading to these results has been sup-

ported by European Union funds and National funds

(GSRT) from Greece and EU under the project JA-

SON: Joint synergistic and integrated use of eArth

obServation, navigatiOn and commuNication tech-

nologies for enhanced border security funded under

the cooperation framework. The contribution of Ms.

Elisavet K. Stathopoulou has been supported by the

European Unions Seventh Framework Programme for

research, technological development and demonstra-

tion under grant agreement no 608013, titled ”ITN-

DCH: Initial Training Network for Digital Cultural

Heritage: Projecting our Past to the Future.”

REFERENCES

Agraﬁotis, P., Doulamis, A., Doulamis, N., and Georgopou-

los, A. (2014a). Multi-sensor target detection and

tracking system for sea ground borders surveillance.

In Proceedings of the 7th International Conference on

PErvasive Technologies Related to Assistive Environ-

ments, page 41. ACM.

Agraﬁotis, P., Georgopoulos, A., Doulamis, A. D., and

Doulamis, N. D. (2014b). Precise 3d measurements

for tracked objects from synchronized stereo-video se-

quences. In Advances in Visual Computing, pages

757–769. Springer.

Alahi, A., Vandergheynst, P., Bierlaire, M., and Kunt, M.

(2010). Cascade of descriptors to detect and track ob-

jects across any network of cameras. Computer Vision

and Image Understanding, 114(6):624–640.

Avidan, S. (2004). Support vector tracking. Pattern Anal-

ysis and Machine Intelligence, IEEE Transactions on,

26(8):1064–1072.

Bay, H., Ess, A., Tuytelaars, T., and Van Gool, L. (2008).

Speeded-up robust features (surf). Computer vision

and image understanding, 110(3):346–359.

Calonder, M., Lepetit, V., Strecha, C., and Fua, P. (2010).

Brief: Binary robust independent elementary fea-

tures. In Computer Vision–ECCV 2010, pages 778–

792. Springer.

Cehovin, L., Kristan, M., and Leonardis, A. (2011). An

adaptive coupled-layer visual model for robust visual

tracking. In Computer Vision (ICCV), 2011 IEEE In-

ternational Conference on, pages 1363–1370. IEEE.

Chermak, L. and Aouf, N. (2012). Enhanced feature de-

tection and matching under extreme illumination con-

ditions with a hdr imaging sensor. In Cybernetic In-

telligent Systems (CIS), 2012 IEEE 11th International

Conference on, pages 64–69. IEEE.

Chermak, L., Aouf, N., and Richardson, M. (2014). HDR

imaging for feature tracking in challenging visibility

scenes. Kybernetes, 43(8):1129–1149.

Collins, R. T., Liu, Y., and Leordeanu, M. (2005). On-

line selection of discriminative tracking features. Pat-

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

628

tern Analysis and Machine Intelligence, IEEE Trans-

actions on, 27(10):1631–1643.

Cui, Y., Pagani, A., and Stricker, D. (2011). Robust point

matching in hdri through estimation of illumination

distribution. In Mester, R. and Felsberg, M., editors,

DAGM-Symposium, volume 6835 of Lecture Notes in

Computer Science, pages 226–235. Springer.

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

dients for human detection. In Computer Vision and

Pattern Recognition, 2005. CVPR 2005. IEEE Com-

puter Society Conference on, volume 1, pages 886–

893. IEEE.

Debevec, P. E. and Malik, J. (1997). Recovering high dy-

namic range radiance maps from photographs. In Pro-

ceedings of the 24th Annual Conference on Computer

Graphics and Interactive Techniques, SIGGRAPH

’97, pages 369–378, New York, NY, USA. ACM

Press/Addison-Wesley Publishing Co.

Doulamis, A., Doulamis, N., Ntalianis, K., and Kollias,

S. (2003). An efﬁcient fully unsupervised video ob-

ject segmentation scheme using an adaptive neural-

network classiﬁer architecture. Neural Networks,

IEEE Transactions on, 14(3):616–630.

Doulamis, N. (2010). Iterative motion estimation con-

strained by time and shape for detecting persons’ falls.

In Proceedings of the 3rd International Conference on

PErvasive Technologies Related to Assistive Environ-

ments, page 62. ACM.

Drago, F., Myszkowski, K., Annen, T., and Chiba, N.

(2003). Adaptive logarithmic mapping for display-

ing high contrast scenes. Computer Graphics Forum,

22:419–426.

Durand, F. and Dorsey, J. (2002). Fast bilateral ﬁltering

for the display of high-dynamic-range images. ACM

Trans. Graph., 21(3):257–266.

Fattal, R., Agrawala, M., and Rusinkiewicz, S. (2007). Mul-

tiscale shape and detail enhancement from multi-light

image collections. ACM Transactions on Graphics

(Proc. SIGGRAPH), 26(3).

Grabner, H., Leistner, C., and Bischof, H. (2008).

Semi-supervised on-line boosting for robust track-

ing. In Computer Vision–ECCV 2008, pages 234–247.

Springer.

Harris, C. and Stephens, M. (1988). A combined corner and

edge detector. In Alvey vision conference, volume 15,

page 50. Manchester, UK.

Henriques, J. F., Caseiro, R., and Batista, J. (2011). Glob-

ally optimal solution to multi-object tracking with

merged measurements. In Computer Vision (ICCV),

2011 IEEE International Conference on, pages 2470–

2477. IEEE.

Huang, C., Wu, B., and Nevatia, R. (2008). Robust ob-

ject tracking by hierarchical association of detection

responses. In Computer Vision–ECCV 2008, pages

788–801. Springer.

Jiang, Z., Huynh, D. Q., Moran, W., Challa, S., and Spadac-

cini, N. (2010). Multiple pedestrian tracking using

colour and motion models. In Digital Image Com-

puting: Techniques and Applications (DICTA), 2010

International Conference on, pages 328–334. IEEE.

Kaaniche, M. B. and Bremond, F. (2009). Tracking hog de-

scriptors for gesture recognition. In Advanced Video

and Signal Based Surveillance, 2009. AVSS’09. Sixth

IEEE International Conference on, pages 140–145.

IEEE.

Kokkinos, M., Doulamis, N. D., and Doulamis, A. D.

(2013). Local geometrically enriched mixtures for sta-

ble and robust human tracking in detecting falls. Int J

Adv Robotic Sy, 10(72).

Kosmopoulos, D. I., Doulamis, N. D., and Voulodimos,

A. S. (2012). Bayesian ﬁlter based behavior recog-

nition in workﬂows allowing for user feedback. Com-

puter Vision and Image Understanding, 116(3):422–

434.

Ladas, N., Chrysanthou, Y., and Loscos, C. Improv-

ing tracking accuracy using illumination neutraliza-

tion and high dynamic range imaging.

Lepetit, V., Lagger, P., and Fua, P. (2005). Randomized

trees for real-time keypoint recognition. In Computer

Vision and Pattern Recognition, 2005. CVPR 2005.

IEEE Computer Society Conference on, volume 2,

pages 775–781. IEEE.

Liu, C., Yuen, J., and Torralba, A. (2011). Sift ﬂow:

Dense correspondence across scenes and its appli-

cations. Pattern Analysis and Machine Intelligence,

IEEE Transactions on, 33(5):978–994.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International journal of computer

vision, 60(2):91–110.

Lucas, B. D., Kanade, T., et al. (1981). An iterative image

registration technique with an application to stereo vi-

sion. In IJCAI, volume 81, pages 674–679.

Mantiuk, R., Myszkowski, K., and Seidel, H.-P. (2006). A

perceptual framework for contrast processing of high

dynamic range images. ACM Trans. Appl. Percept.,

3(3):286–308.

Matthews, I., Ishikawa, T., and Baker, S. (2004). The tem-

plate update problem. IEEE transactions on pattern

analysis and machine intelligence, 26(6):810–815.

Miao, Q., Wang, G., Shi, C., Lin, X., and Ruan, Z. (2011).

A new framework for on-line object tracking based on

surf. Pattern Recognition Letters, 32(13):1564–1571.

Reinhard, E., Stark, M., Shirley, P., and Ferwerda, J. (2002).

Photographic tone reproduction for digital images. In

ACM Transactions on Graphics (TOG), volume 21,

pages 267–276. ACM.

Reinhard, E., Ward, G., Pattanaik, S., and Debevec, P.

(2005). High Dynamic Range Imaging: Acquisi-

tion, Display, and Image-Based Lighting (The Mor-

gan Kaufmann Series in Computer Graphics). Mor-

gan Kaufmann Publishers Inc., San Francisco, CA,

USA.

Robertson, M. A., Borman, S., and Stevenson, R. L. (1999).

Dynamic range improvement through multiple expo-

sures. In In Proc. of the Int. Conf. on Image Processing

(ICIP99, pages 159–163. IEEE.

Rosten, E. and Drummond, T. (2006). Machine learning

for high-speed corner detection. In Computer Vision–

ECCV 2006, pages 430–443. Springer.

HDRImagingforEnchancingPeopleDetectionandTrackinginIndoorEnvironments

629

Rovid, A.and Varkonyi-Koczy, A., Hashimoto, T., Balogh,

S., and Shimodaira, Y. (2007). Gradient based syn-

thesized multiple exposure time hdr image. In Instru-

mentation and Measurement Technology Conference

Proceedings, IMTC 2007, pages 1–6.

Rublee, E., Rabaud, V., Konolige, K., and Bradski, G.

(2011). Orb: an efﬁcient alternative to sift or surf.

In Computer Vision (ICCV), 2011 IEEE International

Conference on, pages 2564–2571. IEEE.

Shi, J. and Tomasi, C. (1994). Good features to track.

In Computer Vision and Pattern Recognition, 1994.

Proceedings CVPR’94., 1994 IEEE Computer Society

Conference on, pages 593–600. IEEE.

Sloan, P.-P., Kautz, J., and Snyder, J. (2002). Precomputed

radiance transfer for real-time rendering in dynamic,

low-frequency lighting environments. In ACM Trans-

actions on Graphics (TOG), volume 21, pages 527–

536. ACM.

Stalder, S., Grabner, H., and Van Gool, L. (2009). Beyond

semi-supervised tracking: Tracking should be as sim-

ple as detection, but not simpler than recognition. In

Computer Vision Workshops (ICCV Workshops), 2009

IEEE 12th International Conference on, pages 1409–

1416. IEEE.

Tang, F., Brennan, S., Zhao, Q., and Tao, H. (2007). Co-

tracking using semi-supervised support vector ma-

chines. In Computer Vision, 2007. ICCV 2007. IEEE

11th International Conference on, pages 1–8. IEEE.

Voulodimos, A. S., Doulamis, N. D., Kosmopoulos, D. I.,

and Varvarigou, T. A. (2012). Improving multi-

camera activity recognition by employing neural net-

work based readjustment. Applied Artiﬁcial Intelli-

gence, 26(1-2):97–118.

Voulodimos, A. S., Kosmopoulos, D. I., Doulamis, N. D.,

and Varvarigou, T. A. (2014). A top-down event-

driven approach for concurrent activity recognition.

Multimedia Tools and Applications, 69(2):293–311.

Yu, Q., Dinh, T. B., and Medioni, G. (2008). Online track-

ing and reacquisition using co-trained generative and

discriminative trackers. In Computer Vision–ECCV

2008, pages 678–691. Springer.

Zhang, L., Li, Y., and Nevatia, R. (2008). Global data asso-

ciation for multi-object tracking using network ﬂows.

In Computer Vision and Pattern Recognition, 2008.

CVPR 2008. IEEE Conference on, pages 1–8. IEEE.

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

630