Using the Scene to Calibrate the Camera

Abstract

Surveillance cameras are used in public and private security systems. Typical systems may contain a large number of different cameras, which are installed in different locations. Manual calibration of each single camera in the network becomes an exhausting task. Although we can find methods that semi-automatically calibrate a static camera, to the best of our knowledge, there is not a fully automatic calibration procedure, so far. To fill this gap, we propose here a novel framework for completely auto-calibration of static surveillance cameras, based on information of the scene (environment and walkers). Characteristics of the method include robustness to walkers' pose and to camera location (pitch, roll, yaw and height), and rapid camera parameter convergence. For a thorough evaluation of the proposed method, the walkers' foot-head projection, the length of the lines projected on the ground plane and the walkers' real heights were analyzed over public and private data sets, demonstrating the potential of the proposed method.

Full text

Using the scene to calibrate the camera
Tiago Trocoli, Luciano Oliveira
Federal University of Bahia (UFBA)
iVision Lab
Salvador, Brazil
http://ivisionlab.dcc.ufba.br
Abstract—Surveillance cameras are used in public and private
security systems. Typical systems may contain a large number
of different cameras, which are installed in different locations.
Manual calibration of each single camera in the network becomes
an exhausting task. Although we can find methods that semi-
automatically calibrate a static camera, to the best of our
knowledge, there is not a fully automatic calibration procedure,
so far. To fill this gap, we propose here a novel framework
for completely auto-calibration of static surveillance cameras,
based on information of the scene (environment and walkers).
Characteristics of the method include robustness to walkers’ pose
and to camera location (pitch, roll, yaw and height), and rapid
camera parameter convergence. For a thorough evaluation of the
proposed method, the walkers’ foot-head projection, the length
of the lines projected on the ground plane and the walkers’
real heights were analyzed over public and private data sets,
demonstrating the potential of the proposed method.
Keywords-camera calibration; surveillance camera; auto cali-
bration;
I. INTRODUCTION
Camera calibration allows mapping of world coordinates
into images, offering several benefits to other Computer Vision
fields. For example, it is possible to decrease the computational
effort of image object search [1], object tracking and pose
estimation can be improved by 3D scene mapping [2], and 3D
information of the scene can provide contextual information
for people re-identification systems [3].
Research in camera auto-calibration becomes of paramount
importance as the use of surveillance camera grows. Existing
methods commonly require constrained assumptions of the
scene to have a camera calibrated. Lv et al. [4] used head-foot
homology to calibrate a camera, demanding a controlled scene
with a tracked and previously known pedestrian in the scene.
Krahnstoever et al. [5] addressed the camera auto-calibration
problem by limited pedestrians detection, and then applying
a bayesian network to find the necessary camera calibration
parameters. Lv et al. [6] extracted two line segments from
the ground plane, specially indicated by the user to detect
vanishing points, in a controlled scene; a person reference
must be selected only when legs cross during human walk in
order to minimize the error in the measurement of a walker’s
major axis. Patwardhan et al. [7] applied a person detection to
find the walkers’ major axes; this constraint requires walkers
in vertical orientation, in the scene. In a controlled indoor
scene, Micusik et al. [8] used shape object detection to
estimate walkers’ major axes and the distance between objects
Fig. 1. Outline of the proposed auto-calibration framework. Given a sequence
of images, an adaptive background subtraction extracts the background and
walkers’ blobs. Manhattan World based vanishing points are detected in the
background image, and then used in a convergence filter to select the blobs,
which point to vertical vanishing point. Walker’s relative height are computed
to fit a height distribution, and, finally, the required camera calibration
parameters are estimated.
and the camera; that turns the process to have an intensive
computational effort, limiting the possible human poses to
be searched in the images. Liu et al. [9] used a walker’s
major axis line, extracted by a background subtraction method,
to calibrate a surveillance camera via a Bayesian approach;
a large set of samples was required to achieve reasonable
precision. Lee et al. [10] exploited scene clues for detection
of vanishing points, requiring an object size as reference to

Fig. 2. Camera model and representation of world coordinate system.
work and a high level of human intervention.
A. Contributions
This paper introduces a method which can overcome the
main limitations of the existing approaches, mainly regarding
the constraints demanded for the other methods. Basic feature
extraction from walkers and scenario structures is made fully
automatic, not demanding any human intervention, which is
required in almost all related works [4], [10]. Our method
is robust to variation of walkers’ pose, and camera position
and orientation (pitch, roll, yaw and height), avoiding the use
of person detectors (and, consequently, avoid failing when
detection fails), as found in [7], [8], [5], [6]; this leads our
method to have more flexibility and accuracy. The extraction
of environment structure cues helps reducing the amount of
samples necessary for camera calibration, in contrast to [9],
which does not exploit this type of information. Table I
summarizes the comparison at a glance.
The main goal of our method is to calibrate a pinhole-based
camera fully automatic (see Sec. II). The main assumptions
of the proposed method are twofold: (i) based on the Man-
hattan World constraint and (ii) considering prior information
on walkers, walkers’ major axes are first computed, and
thereafter filtered by a vertical vanishing point convergence
filter – as described in Sec. III. Scene structure features,
computed against the image background, support vanishing
point detection, which is addressed in Sec. IV. Relative height
and height distribution are computed to fit the average of a
walkers’ relative height of the local distribution – described
in Sec. V. After defining all required calibration parameters,
the performance of the proposed method is assessed in Sec.
VI. Sec VII draws the conclusions. All these main steps of the
auto-calibration framework are depicted in Fig. 1.
II. CAMERA MO DE L
Here the camera model is assumed to be pinhole. This
assumption can be used for all types of static surveillance
camera, without loss of generality. Thus, the relationship
between the 3D point P= [X, Y, Z, 1]Tand its image
projection p= [u, v, 1]Tis given by
p=K[R|t]P(1)
where Kis the intrinsic matrix, and [R|t]represents the extrin-
sic matrix, composed of a rotation matrix R=RZRXRYand
a translation vector t= [0,0,−hc]T. As in [11], we assume
zero-skew, aspect ratio equal to one and image principal point
center (0,0), so that the focal length fis the only parameter
to be found in K. In the world coordinate system (WCS), the
origin point is placed on the ground plane as a projection of
the camera center point. These assumptions narrow the camera
calibration problem down to three required parameters (θ,ρ,
hc), which are needed besides f, to map a point in WCS to
the camera coordinate system; θand ρrepresent the camera
rotation around the X-axis (RX(θ)) and the Z-axis (RZ(ρ))
in matrix R, respectively, and hcdenotes the translation (t)
in Z-axis orientation. θis bounded by π/2(camera is parallel
to the ground plane) and π(camera is pointing directly to
ground plane). The yaw camera orientation does not contribute
to camera calibration when the ground plane is taken as the
reference. This assumption allows to set RYas a 3X3identity
matrix [4].
Liebowitz and Zisserman [11] show that tree orthogonal
vanishing points are directly associated with Kand Rso that it
is possible to recover f,θand ρfrom only image information.
So, by representing the vanishing points in image coordinates,
as vpi= (ui, vi), where vp0is the vertical vanishing point,
and vp1and vp2are the vanishing points of the horizon lines,
fcan be calculated as
f=pd(center, vp0)∗l(2)
where d(., .)is the distance between two points, and lis the
distance between the camera center point and the horizon line;
now, θand ρcan be defined as
θ=atan2qu2
0+v2
0,−f(3)
ρ=atan(−u0/v0)(4)
The last parameter to be found is the camera height, hc. For
that, the walkers’ relative heights, hi, achieved by the ratio of
walkers’ real height, h3D
i, and hc, are defined as
hi=h3D
i
hc
= 1 −d(ph, ql)d(pf, vp0)
d(pf, ql)d(ph, vp0)(5)
where phand pfare the highest and the lowest point,
respectively, of walkers’ major axes, and qlis the intersection
point between the horizon line and the line defined by phand
pf.
According to Criminisi, Reid and Zisserman [12], Eq. (5)
provides a way to determine hiby means of the horizon
line and the vertical vanishing point vp0. Each walkers’
relative height keeps its value regardless of walker’s major
axis variation (e.g., due to perspective projection). This way,
it is possible to apply the prior distribution of the walkers’ real
heights, h3D
i, on hi, with no need of knowing the real height
of any walker in the monitored scene. By rewriting Eq. (5), it
is possible to estimate hcas the ratio between the average of
the walkers’ real height, H3D, and the expectation, E, of hi,
given by

TABLE I
COMPARATIVE SUMMARY OF THE WORKS ON CAMERA CALIBRATION.
Ref Walkers Features Environment Features Reference Measure
Manual Auto Manual Auto Manual Auto
[4]
Background
subtraction
with manual
axis
Track a walker
with known
height
[5]
Limited walker
detection,
and specific
to the scene
Require exactly
average of walkers
height
[6]
Background
subtraction
with cross leg
detection
Track a walker
with known height
[7]
Apply a walker
detector with
camera position
constraints
Track a walker
with known height
[8]
Use specific
shape person
detector for
camera position
constraint
Know the person
within controlled
environment
[9]
Background
subtraction with
main axis
estimator
Uses the human
height distribution
as reference
[10] Walker main axis
inserted manually
4 lines marked on
background scene,
and each pair should
point for differents
vanishing points
Require a known
height object
Ours
Based on
completely
automatic
background
subtraction
Completely automatic
vanishing point detection,
and scene line detection
Use height
distribution
and scene cues
as reference
hc=H3D
E(hi)(6)
Figure 2 illustrates the geometric representation of the
camera model and its parameters.
III. BACKGRO UN D SUBTRACTION
Methods which use specialized person detection to estimate
camera parameters limit the calibration process to a particular
camera position and orientation [7], [8]. Instead, the goal here
is to use background subtraction (BS) to provide a flexible
way of extracting the walkers’ major axes, in an unsuper-
vised manner. This brings the advantage of not restricting
the calibration method to a previously known scene. On the
other hand, if the calibration method solely relies on blob
extraction for estimation of camera parameters, this can fail if
BS fails. Warped blobs tend to induce erroneous major axis
orientation and size1. This is because the major axes do not
intersect the vertical vanishing point anywhere (see Fig. 5(a)).
To circumvent this orientation problem, after estimating the
feet-head homology, (pfph), and the vanishing points (see Sec.
IV), a convergence filter is applied in the major axes, achieved
1Here, blob shape is approximated by an ellipsis.
Fig. 3. Comparative performance of background subtraction methods.
by the foreground blobs, in order to select those major axes
that point to vp0(see Fig. 5(b)). To overcome the size problem,
a RANSAC is used to select the major axes, which fit a height
distribution; see Sec. V and Fig. 5(c), for more details.
Broadly speaking, the characteristics of a BS method to

Fig. 4. Example of dilate and erode operators applied on walker blob.
accomplish our task should include speed and segmentation
accuracy. Here we propose a modified version of the adaptive
learning method, with the addition of a dilate and an erode
morphological operations. The goal of using the morphological
operations was to reduce common failures (present in the
baseline method), and to preserve the original walker height,
as illustrated in Fig. 4. Morphological operations also cause
orientation and size distortions in the walkers’ major axes.
Also, instead of comparing pixels in gray level, moving objects
and background are compared across the three RGB image
channels.
To assess the performance of our method, a special data set
with 2000 frames, gathered from PETS 2006 [13], was used
with real surveillance scenes, four different camera positions
and orientations, (including walkers’ occlusion) and adaptive
background. Only moving objects were considered to be
annotated as the ground truth. The best four public available
methods evaluated in [14] were compared against ours: Stan-
dard adaptive background learning, mixture of gaussian [15],
multi-layer [16] and adaptive SOM [17].
Two metrics were considered to compare the methods:
frames per seconds (FPS) and the area under receiver op-
erating characteristic curve (AUC) as shown in Fig. 3. The
standard adaptive background learning was the fastest among
all, however with the worst AUC. Multi-layer was the slowest
one. Our method presented the best AUC, and it was near
the fastest method, becoming the most feasible method to be
applied in our approach.
IV. VANISHING POI NT DETECTION
After suppressing moving objects with the BS method, line
segments are found by a modified version of the line segment
detector (LSD) [18], as illustrated in Figs. 6 (Left column),
at different camera locations. These line segments are the
scene structure clues, and they were extracted from static
structures presented in the background image. This bundle
of line segments allows the use of a reliable and specialized
method to estimating vanishing points. Based on Manhattan
World assumption, these lines segments point to, at least,
3 vanishing points, which is enough for camera calibration.
This approach does not need major axes in the estimation
of vanishing points. This increases vanishing point accuracy,
allowing for the use of few samples for calibration parameter
convergence.
After line segment detection, a RANSAC is used to find the
vanishing points with four examples of line segments, which is
assumed to follow a Manhattan World assumption. Prior to the
RANSAC computation, our method makes an oriented search
for the samples used in RANSAC. During the line segment
extraction, a histogram of line segment orientations (between
0◦and 180◦) is built, searching for convergence peaks. Line
segments inside the regions of the peaks in the histogram
demand few RANSAC iterations to converge to the vanishing
points.
Although there are other methods to detect the vanishing
points from line segments, our method demonstrated to be
the fastest found, detecting enough vanishing points necessary
for camera calibration, as well as, avoiding analyzing all
points; some results are shown in Figs. 6(Right column).
Vanishing points allow defining some camera parameters (f,
θand ρ), and help selecting the walkers’ major axes by using
the convergence filter (good walkers). This latter chooses line
segments by means of Liebowitz’s distance [11] to the vertical
vanishing points, as will be described further.
The convergence filter selects the principal axis which point
to the vertical vanishing point, restricting blobs with wrong
orientation. Yet simple, the convergence filter eliminates ap-
proximately 90% of the noise presented in the set of principal
axis. This effect is illustrated in Figs. 5(a) and (b). With a set
of more consistent and smaller amount of data, the reference
height procedure (see Sec. ) requires less iteration during the
selection of the best sampling set.
V. WALK ER S’ HEIGHT DISTRIBUTION
After determining vanishing points, focal length, f, was
found according Eq. 2, while ρand θwere determined by
Eqs. 4 and 3, respectively. So far, the only missing parameter
is the camera height. To keep the premisse to provide a fully
automatic camera calibration framework, the camera height,
hc, is estimated by approximating a height distribution relative
to real height distribution. According to Liu et al. [9], human
real height distribution encompasses 90% of the heights, h3D
i,
around the average height, H3D. Yet, the set of human real
heights has a relative difference of less than 7.6% from the
mean. This prior distribution allows matching the average
human height to the mean of the walkers’ relative heights (as
shown in Eq. (6)), and it is given by
|hi−µ|
µ≤0.076 (7)
where µis the mean value of the relative heights.
To fit the relative height to the major’ axes found, a
RANSAC is used to find the best samples which define the
distribution. Randomly, RANSAC selects samples to compute
average relative height. After that, Eq. (7) is applied to
each sample. The sample is discarded in case of overcoming
the boundary defined in Eq. (7). This process is repeated
iteratively, and the biggest sample is returned, at the end. After
that, the mean, µ, of the relative heights its standard deviation

Fig. 5. (a) All extracted major axes; (b) major axes after the convergence filter; (c) selected walkers’ major axes which fit to a human height distribution.
Fig. 6. Examples of vanishing point detection. Left column shows detections of line segments with LSD method [18]; right column show three detected
vanishing points.
are calculated. The heights form a normal distribution, which
allow to restrict the search to 95% of the elements in 2σ
from the mean. This constraint is applied to 90% of the
elements (relative heights), providing the selection of the
relative walkers’ heights by means of the distance between
each height and the mean. To each RANSAC iteration, a mean
value of the walkers’ relative height is assigned considering
the interval [µ−2σ, µ + 2σ].
VI. EX PE RI ME NTAL EVAL UATION
Evaluation of the proposed method was assessed from four
different data sets: PETS 2006 [13], PETS 2007 [19], CVLAB
[20], and a private gathered data set. The goal was to evaluate
the value of the relative foot-head homology root mean square
error (FHH RMSE), ground-plane measurements (determined
by the distance between two points in the ground plane) and
walkers’ real heights. Each data set provided a way to analyse
those metrics, as showed in Table II, which summarizes the
characteristics of each one of the data sets, regarding the type
of environment (indoor/outdoor), occupation of the scene by
the people (partially crowded, crowded or not crowded), how
the scene was captured concerning the randomness of peo-
ple walking (uncontrolled, partially controlled or controlled),
number of views and frames evaluated, and the applied type
of error analysis. Additionally, the number of necessary good
walkers was computed in order to assess how fast the method
find the camera parameters. To the best of our knowledge is
the first time that a thorough performance assessment of this
kind is made.
Figure 7 shows the evaluation of the heights and the distance
of points on the ground plane with respect to the number of
samples tested, over three different camera heights (close to
the floor, around 2 meters, and around 4 meters). Each video
takes less than 1 minute to process the scene at 25 FPS, and
all walkers’ heights are previously known. Height evaluation

TABLE II
CHA RACT ER IST IC S OF TH E DATA SET S USE D
Dataset Enviroment Conditions Scene # of views # of frames Type of analysis
PETS 2006 indoor partially
crowded uncontrolled 3 2500 ground-plane measurements
FHH RMSE
PETS 2007 indoor crowded uncontrolled 2 2500 FHH RMSE
CVLAB indoor, outdoor not crowded partially
controlled 5 2500 FHH RMSE
Ours indoor not crowded controlled 3 2500 real height
Fig. 7. Cumulative curves of (from left to right): heights, ground-plane measurements and foot-head homology (FHH) RMSE. Left plot: σrepresents an
upper bound of a relative error found on a percentage of samples tested in the data set; center plot: λrepresents an upper bound of an absolute error (in
meters) found on the percentage of samples tested in the data set; right plot: δrepresents an upper bound of a relative error found on FHH RMSE of samples
tested on PETS 2006, PETS 2007 and CVLAB data sets. Our gathered and PETS 2006 data sets were used to evaluate the real heights and ground-plane
measurements, respectively.
Fig. 8. Examples of camera auto-calibration results. Green line segments represent the ground truth manually labeled; Red line segments represent the
estimated feet-head projection; Cyan circles represent the person position at the ground plane. The first row contains images from PETS 2006 (the first three)
and from PETS 2007 (the last two). Second row contains images from CVLAB data set.
demonstrated that our method is able to have an absolute
error below 5 cm in 70% of the samples tested (see Fig. 7
(center)). Yet, if one considers, for example, the same 70% of
the samples tested, our method presents only 3% of relative
error (see Fig. 7 (left)).
The ground truth landmarks in PETS 2006 data sets show
the distances between several pair of points on the ground
plane. It was observed that distances closer to the camera
are more accurate, due to the camera perspective distortion
(see Fig. 9). Despite this issue, our method was able to
have a maximum relative error of 15% among the estimated
distance on the ground (see Fig. 7 (left)). Figure 8 shows
visual examples of the achieved results on the PETS 2006
and CVLAB data sets.
The relative FFH RMSE provides a way to evaluate camera
calibration accuracy from perspective of image projection,
numerically. Our method was submitted to different point of
views in CVLAB data sets, and achieved an FHH RMSE of
0.03, overcoming the method proposed by [9], as shown in
Table III. These results were achieved even with our proposed
method exposed to varying lighting conditions in indoor and
outdoor environments. PETS 2006 and 2007 portray data sets
with uncontrolled scene and crowded occupation. These data
sets showed the method is sensitive to crowd occupation,
due the walkers’ major axis distortion, which occurred in the
foreground segmentation. However, FHH RMSE evaluation

Fig. 9. Ground plane estimation samples for distinct views from PETS 2016.
The blue and red lines portray the ground plane, and yellow line segments
represent are ground plane perpendicular.
TABLE III
COMPARATIVE EVALUATION OF THE CALIBRATION PROCESS ON THE
CVLAB DATA SET.
Metrics Liu et al [9] Our Method
Average of good walkers
necessary for convergence 1800 370
FHH RMSE 0.05 0.03
reaches 0.07 as a maximum error, indicating that the method
is also robust to handle with crowded scene. Table III shows
also the comparison of our method against [9] with respect
to the average number of good walkers necessary to camera
parameter convergence; in this case, our method requires
almost 5 times less examples than the method proposed in
[9]. Figure 8 shows visual examples of the achieved results
on the PETS 2006, PETS 2007, CVLAB data sets.
Our method can handle with a low resolution representation
of a person, as shown in Fig. 6 (first row and column), due
to the BS method sensitivity. The proposed framework can be
applied in several indoor public spaces (e.g. airports, malls,
buildings), and outdoor places, mainly occupied by people
and with some man-made structures in the background (e.g.
squares, parks).
VII. CONCLUSION
In this paper, a novel framework for fully automatic calibra-
tion of static surveillance cameras based on scene cues was
presented. The proposed method demonstrated to have high
accuracy, very fast calibration parameter convergence and no
need of human intervention. Although BS methods usually
produce a lot of noise in blob extraction – used to define
walkers’ major axes – the proposed camera calibration frame-
work is able to overcome this problem, still demonstrating high
accuracy at the end. The prior information on walkers, along
with the walkers’ relative heights and scene cues, avoid the
need of knowing any reference length in the scene. As future
work, we are investigating a method to compute automatically
the radial distortion of the camera lens.
REFERENCES
[1] S. Rujikietgumjorn and R. Collins, “Optimized pedestrian detection for
multiple and occluded people,” in IEEE Conference on Computer Vision
and Pattern Recognition, 2013, pp. 3690–3697.
[2] A. Balan, L. Sigal, and M. Black, “A quantitative evaluation of video-
based 3d person tracking,” in IEEE International Workshop on Visual
Surveillance and Performance Evaluation of Tracking and Surveillance,
2005, pp. 349–356.
[3] Z. Wu, Y. Li, and R. Radke, “Viewpoint invariant human re-identification
in camera networks using pose priors and subject-discriminative fea-
tures,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
pp. 1095–1108, 2015.
[4] F. Lv, Z. T., and R. Nevatia, “Self-calibration of a camera from video
of a walking human,” in IEEE International Conference on Pattern
Recognition, 2002, pp. 562–567 vol.1.
[5] N. Krahnstoever and P. Mendonca, “Bayesian autocalibration for surveil-
lance,” in IEEE International Conference on Computer Vision, 2005, pp.
1858–1865.
[6] F. Lv, T. Zhao, and R. Nevatia, “Camera calibration from video of a
walking human,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, pp. 1513–1518, 2006.
[7] D. Rother, K. Patwardhan, and G. Sapiro, “What can casual walkers tell
us about a 3d scene?” in IEEE International Conference on Computer
Vision, 2007, pp. 1–8.
[8] B. Micusik and T. Pajdla, “Simultaneous surveillance camera calibration
and foot-head homology estimation from human detections,” in IEEE
Conference on Computer Vision and Pattern Recognition, 2010, pp.
1562–1569.
[9] J. Liu, R. Collins, and Y. Liu, “Surveillance camera autocalibration
based on pedestrian height distributions,” in British Machine Vision
Conference, 2011, pp. 144–154.
[10] S. Lee and R. Nevatia, “Robust camera calibration tool for video
surveillance camera in urban environment,” in IEEE Computer Society
Conference on Computer Vision and Pattern Recognition Workshops,
2011, pp. 62–67.
[11] D. Liebowitz and A. Zisserman, “Combining scene and auto-calibration
constraints,” in IEEE International Conference on Computer Vision.,
1999, pp. 293–300.
[12] A. Criminisi, I. Reid, and A. Zisserman, “Single view metrology,”
International Journal of Computer Vision, vol. 40, no. 2, pp. 123–148,
2000.
[13] I. C. on Computer Vision and P. Recognition, “Benchmark data,” 2006,
available from http://www.cvg.reading.ac.uk/PETS2006/data.html.
[14] A. Sobral and A. Vacavant, “A comprehensive review of background
subtraction algorithms evaluated with synthetic and real videos,” Com-
puter Vision and Image Understanding, vol. 122, pp. 4–21, 2014.
[15] P. KaewTraKulPong and R. Bowden, “An improved adaptive background
mixture model for real-time tracking with shadow detection,” in Video-
Based Surveillance Systems. Springer US, 2002, pp. 135–144.
[16] J. Yao and J. Odobez, “Multi-layer background subtraction based on
color and texture,” in IEEE Conference on Computer Vision and Pattern
Recognition, 2007, pp. 1–8.
[17] L. Maddalena and A. Petrosino, “A self-organizing approach to back-
ground subtraction for visual surveillance applications,” IEEE Transac-
tions on Image Processing, pp. 1168–1177, 2008.
[18] R. Gioi, J. Jakubowicz, J. Morel, and G. Randall, “Lsd: A fast line
segment detector with a false detection control,” IEEE Transactions on
Pattern Analysis and Machine Intelligence, pp. 722–732, 2010.
[19] I. C. on Computer Vision and P. Recognition, “Benchmark data,” 2007,
available from http://www.cvg.reading.ac.uk/PETS2007/data.html.
[20] C. V. Laboratory, “Multi-camera pedestrians video,” 2013, available
from http://cvlab.epfl.ch/data/pom/.