RECTANGULAR EMPTY PARKING SPACE DETECTION USING
SIFT BASED CLASSIFICATION
Harish Bhaskar, Naoufel Werghi
Dept. of Computer Engineering, Khalifa University, Sharjah Campus, Sharjah, U.A.E.
Saeed Al-Mansoori
Dept. of Communication Engineering, Khalifa University, Sharjah Campus, Sharjah, U.A.E.
Keywords:
Radon transform, Parking space detection, SIFT, Supervised classification, Peak extraction, Filtering, Geo-
metric and spatial constraints.
Abstract:
In this paper, we describe a method of combining rectangle detection and scale invariant feature transform
(SIFT) analysis for empty parking space detection. A parking space in a parking lot is represented as a rectan-
gular region of pixels in an image captured from an aerial camera. Detecting rectangular parking spaces in a
new image involves an alternating scheme of extracting peaks from the Radon transform for the whole image
and filtering them against specific geometric and spatial constraints. We then compute SIFT descriptors from
these detected rectangular parking spaces and further apply supervised classification methods for detecting
empty parking spaces. We demonstrate the performance of our model on several synthetic and real data.
1 INTRODUCTION
With increasing number of vehicles over years, park-
ing has become an important issue particularly in
commercial environments such as shopping malls and
airports. Empty parking space detection system that
can automatically identify empty parking spaces and
guide users to it will save lot of time, money, effort
and are highly desirable. Recently, a number of so-
lutions that employ different sensors have been sug-
gested, but they are either too expensive to implement
or have failed to be effective. According to (Bong
et al., 2008), we can conveniently categorize these
car parking management or guidance system based on
their technologies into: counter-based, wired sensor-
based, wireless sensor-based and image-based.
The simplest is the counter-based system that
counts the number of cars that enter and exit a car
park area (Ristola, 1992). This system is capable
of providing information on the total number empty
parking spaces in an area but cannot guide the driver
to the location of empty parking spaces. The wired
sensor system rely on installing ultrasonic systems in
each parking lot to detect the occupancy of parking
spaces. These sensors are managed by a control unit
to which they are wired and are capable of direct-
ing drivers to empty parking spaces. Wired systems
differ by the type of sensor that is used. Different
sensors including but not restricted to vortex berth
detector (Yu and Liu, 2004), magnetic field sensor
(Wolff et al., 2006) are currently available for parking
guidance. These systems are limited due to the high
installation and maintenance costs that results from
the long and complicated wiring that is required to
get them functional. A simple extension of wireless
sensor systems to parking guidance have also been
available in recent years. The sensors in most wire-
less systems are micro-controlled and typically in-
clude multi-sensor controls. They are more effective
than wired sensor system but are very expensive than
most other systems. Some common examples of these
systems include: (Tang et. al, 2006) that uses the
extended crossbow network architecture, (Benson et
al., 2006) that adopts an anisotropic magneto-resistive
magnetic field sensor together with a microprocessor
transceiver.
Finally, image based techniques based on video
capture sensors are also commonly used. These are
particularly suitable where parking areas are already
monitored by CCTV surveillance systems. Image
based parking guidance systems are simple and in-
expensive. However, transmitting image or video in-
formation through wireless sensors could be slightly
costly. There has been some recent research efforts
214
Bhaskar H., Werghi N. and Al-Mansoori S..
RECTANGULAR EMPTY PARKING SPACE DETECTION USING SIFT BASED CLASSIFICATION.
DOI: 10.5220/0003358702140220
In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2011), pages 214-220
ISBN: 978-989-8425-47-8
Copyright
c
2011 SCITEPRESS (Science and Technology Publications, Lda.)
that has focused on the applications of car parking
system using image sensor technologies as in (Funck
et al., 2004). Other examples include: (Seo et al.,
2009; Wang and Hanson, 1998) where empty park-
ing spaces are estimated using an aerial image and
semi-supervised learning mechanism and (Bong et
al., 2008) where an integrated approach using im-
age processing algorithms called Car Occupancy In-
formation System (COINS) for accurate and robust
monitoring of parking spaces. However, these mod-
els often need manual seeding and are restrictive for
arrangement of parking spaces.
and inexpensive. However, transmitting image or
video information through wireless sensors could be
slightly costly. There has been some recent research
efforts that has focused on the applications of car
parking system using image sensor technologies as
in (Funck et al., 2004). Other examples include: (Seo
et al., 2009; Wang and Hanson, 1998) where empty
parking spaces are estimated using an aerial image
and semi-supervised learning mechanism and (Bong
et al., 2008) where an integrated approach using
image processing algorithms called Car Occupancy
Information System (COINS) for accurate and robust
monitoring of parking spaces. However, these models
often need manual seeding and are restrictive for
arrangement of parking spaces.
Figure 1: Schematic illustration of the proposed empty car
parking space detection technique
In this paper, we explore an unified framework for
detecting rectangular parking space using combined
Radon transform and SIFT descriptor based classifi-
cation for image based empty parking space detec-
tion. Our model combines geometric constraints over
the Radon transform domain with spatial constraints
in the co-ordinate (or image) domain for accurate and
simultaneous detection of multiple rectangular park-
ing spaces. Further, we integrate SIFT descriptor
analysis and classification for efficient empty parking
space detection. An illustration of the framework is
presented in Figure 1.
Contributions & Structure: The method proposed
in this paper combines Radon transform based rectan-
gle detection for parking space extraction with SIFT
analysis (Lowe, 2004) based empty parking space
classification. One novelty of the method is the inte-
grated three stage framework for empty parking space
detection. In the first step, we use Radon transform
based rectangle detection that allows detecting park-
ing spaces in real image and is robust to the presence
of noise and illumination changes. In the second step,
we classify the extracted parking spaces as empty
spaces using SIFT features and supervised classifica-
tion techniques. Our results indicate that using the
proposed framework presents good performance at
high accuracy rates.
2 Method
We formulate our proposed method as a three
stage process of: rectangular parking space extrac-
tion using Radon transform based rectangle detection
algorithm, SIFT feature detection on extracted park-
ing spaces and finally supervised classification. We
represent a parking space in the image as a rectangu-
lar region R. Given an image I of the parking space,
our first step is to extract n parking spaces R
n
from
I. Once, we obtain the rectangular parking space re-
gions, we perform SIFT analysis on each of the n
parking spaces to obtain S
n
, where S
i
is a feature vec-
tor containing descriptors of that region R
i
. These
feature descriptors are further classified into either
empty or full classes representing whether the parking
space is empty or occupied using a supervised classi-
fier. We summarize our approach as follows:
1. Input any given image I of the parking area.
2. Parking space extraction: We adopt a rectangle
detection algorithm based on Radon transform for
parking space extraction. i.e R
n
= Γ(I), where Γ
is the rectangle detection function and R
n
are the
n rectangle regions extracted from I
3. SIFT Feature Detection: For every parking space
region i = 1 ... n,
• Compute SIFT descriptors: S
i
= ψ(R
i
), where
ψ(.) is the SIFT operator
4. Perform binary Classification of descriptors to ob-
tain class label ` = −1or1, where -1 represents
empty parking space and 1 represents an occupied
parking space for each of n SIFT descriptors S
n
.
i.e. ` = χ(S
n
), where χ(.) represents the classifier.
In the following subsections, we describe the
Γ(.), psi(.) and χ(.) functions in more detail. We
formulate the problem of rectangular parking space
detection as a alternating scheme of transform peak
extraction and combined spatial and transform
domain geometric constraints based peak filtering as
depicted in Figure 2.
Figure 2: Schematic illustration of the proposed rectangular
parking space detection methodology
Figure 1: Schematic illustration of the proposed empty car
parking space detection technique.
In this paper, we explore an unified framework for
detecting rectangular parking space using combined
Radon transform and SIFT descriptor based classifi-
cation for image based empty parking space detec-
tion. Our model combines geometric constraints over
the Radon transform domain with spatial constraints
in the co-ordinate (or image) domain for accurate and
simultaneous detection of multiple rectangular park-
ing spaces. Further, we integrate SIFT descriptor
analysis and classification for efficient empty parking
space detection. An illustration of the framework is
presented in Figure 1.
Contributions & Structure: The method proposed
in this paper combines Radon transform based rectan-
gle detection for parking space extraction with SIFT
analysis (Lowe, 2004) based empty parking space
classification. One novelty of the method is the inte-
grated three stage framework for empty parking space
detection. In the first step, we use Radon transform
based rectangle detection that allows detecting park-
ing spaces in real image and is robust to the presence
of noise and illumination changes. In the second step,
we classify the extracted parking spaces as empty
spaces using SIFT features and supervised classifica-
tion techniques. Our results indicate that using the
proposed framework presents good performance at
high accuracy rates.
2 METHOD
We formulate our proposed method as a three stage
process of: rectangular parking space extraction using
Radon transform based rectangle detection algorithm,
SIFT feature detection on extracted parking spaces
and finally supervised classification. We represent a
parking space in the image as a rectangular region R.
Given an image I of the parking space, our first step
is to extract n parking spaces R
n
from I. Once, we
obtain the rectangular parking space regions, we per-
form SIFT analysis on each of the n parking spaces to
obtain S
n
, where S
i
is a feature vector containing de-
scriptors of that region R
i
. These feature descriptors
are further classified into either empty or full classes
representing whether the parking space is empty or
occupied using a supervised classifier. We summarize
our approach as follows:
1. Input any given image I of the parking area.
2. Parking space extraction: We adopt a rectangle
detection algorithm based on Radon transform for
parking space extraction. i.e R
n
= Γ(I), where Γ is
the rectangle detection function and R
n
are the n
rectangle regions extracted from I
3. SIFT Feature Detection: For every parking space
region i = 1 . . . n,
• Compute SIFT descriptors: S
i
= ψ(R
i
), where
ψ(.) is the SIFT operator
4. Perform binary Classification of descriptors to ob-
tain class label ` = −1or1, where -1 represents
empty parking space and 1 represents an occupied
parking space for each of n SIFT descriptors S
n
.
i.e. ` = χ(S
n
), where χ(.) represents the classifier.
In the following subsections, we describe the Γ(.),
psi(.) and χ(.) functions in more detail. We formu-
late the problem of rectangular parking space detec-
tion as a alternating scheme of transform peak extrac-
tion and combined spatial and transform domain geo-
metric constraints based peak filtering as depicted in
Figure 2.
and inexpensive. However, transmitting image or
video information through wireless sensors could be
slightly costly. There has been some recent research
efforts that has focused on the applications of car
parking system using image sensor technologies as
in (Funck et al., 2004). Other examples include: (Seo
et al., 2009; Wang and Hanson, 1998) where empty
parking spaces are estimated using an aerial image
and semi-supervised learning mechanism and (Bong
et al., 2008) where an integrated approach using
image processing algorithms called Car Occupancy
Information System (COINS) for accurate and robust
monitoring of parking spaces. However, these models
often need manual seeding and are restrictive for
arrangement of parking spaces.
Figure 1: Schematic illustration of the proposed empty car
parking space detection technique
In this paper, we explore an unified framework for
detecting rectangular parking space using combined
Radon transform and SIFT descriptor based classifi-
cation for image based empty parking space detec-
tion. Our model combines geometric constraints over
the Radon transform domain with spatial constraints
in the co-ordinate (or image) domain for accurate and
simultaneous detection of multiple rectangular park-
ing spaces. Further, we integrate SIFT descriptor
analysis and classification for efficient empty parking
space detection. An illustration of the framework is
presented in Figure 1.
Contributions & Structure: The method proposed
in this paper combines Radon transform based rectan-
gle detection for parking space extraction with SIFT
analysis (Lowe, 2004) based empty parking space
classification. One novelty of the method is the inte-
grated three stage framework for empty parking space
detection. In the first step, we use Radon transform
based rectangle detection that allows detecting park-
ing spaces in real image and is robust to the presence
of noise and illumination changes. In the second step,
we classify the extracted parking spaces as empty
spaces using SIFT features and supervised classifica-
tion techniques. Our results indicate that using the
proposed framework presents good performance at
high accuracy rates.
2 Method
We formulate our proposed method as a three
stage process of: rectangular parking space extrac-
tion using Radon transform based rectangle detection
algorithm, SIFT feature detection on extracted park-
ing spaces and finally supervised classification. We
represent a parking space in the image as a rectangu-
lar region R. Given an image I of the parking space,
our first step is to extract n parking spaces R
n
from
I. Once, we obtain the rectangular parking space re-
gions, we perform SIFT analysis on each of the n
parking spaces to obtain S
n
, where S
i
is a feature vec-
tor containing descriptors of that region R
i
. These
feature descriptors are further classified into either
empty or full classes representing whether the parking
space is empty or occupied using a supervised classi-
fier. We summarize our approach as follows:
1. Input any given image I of the parking area.
2. Parking space extraction: We adopt a rectangle
detection algorithm based on Radon transform for
parking space extraction. i.e R
n
= Γ(I), where Γ
is the rectangle detection function and R
n
are the
n rectangle regions extracted from I
3. SIFT Feature Detection: For every parking space
region i = 1 .. . n,
• Compute SIFT descriptors: S
i
= ψ(R
i
), where
ψ(.) is the SIFT operator
4. Perform binary Classification of descriptors to ob-
tain class label ` = −1or1, where -1 represents
empty parking space and 1 represents an occupied
parking space for each of n SIFT descriptors S
n
.
i.e. ` = χ(S
n
), where χ(.) represents the classifier.
In the following subsections, we describe the
Γ(.), psi(.) and χ(.) functions in more detail. We
formulate the problem of rectangular parking space
detection as a alternating scheme of transform peak
extraction and combined spatial and transform
domain geometric constraints based peak filtering as
depicted in Figure 2.
Figure 2: Schematic illustration of the proposed rectangular
parking space detection methodology
Figure 2: Schematic illustration of the proposed rectangular
parking space detection methodology.
RECTANGULAR EMPTY PARKING SPACE DETECTION USING SIFT BASED CLASSIFICATION
215
We represent the rectangular parking space as two
pairs of transform peaks that satisfy certain specific
geometric constraints in the transform domain in ad-
dition to the spatial constraints of the corresponding
line segments in spatial co-ordinate (image) domain.
Given an image I, our aim is then to find rectangle(s)
R each represented as two pairs of peaks PP
1
and
PP
2
, where PP represents a Peak Pair that satisfy
certain conditions. We summarize our approach as
follows:
1. Edge Detection and Enhancement: Get the edge
image using: E = (I), where (.) is the edge de-
tection and enhancement operator. We perform
conventional edge detection using the Canny op-
erator and further enhance the edge image using
classical denoising, edge linking and edge clean-
ing operations. We use this edge image E for fur-
ther processing.
2. Peak Extraction: Apply the Radon transform on
the edge image: A = T(E), where T(.) represents
the Radon transform. The Radon transform of an
image pixel I(x, y) as the integral projection along
a straight line defined by its distance ρ from the
origin and its angle of inclination θ, a definition
that is similar to that of the Hough transform. The
Radon transform is considered to be the gener-
alization of the Hough transform and has been
proven to be more robust to noise for rectangle
detection in (Bhaskar et al., 2010). Mathemati-
cally,
A(ρ, θ) =
Z
x
Z
y
Iδ(xcosθ + ysinθ − ρ)dxdy (1)
where the δ is the dirac delta function defines in-
tegration only over the line. The range of θ lim-
ited to [0, π] . Similar to the Hough transform,
each point in the Radon transform space corre-
spond to a straight line in the spatial co-ordinate
domain. Conversely, each point in the spatial do-
main becomes a sine curve in the projection do-
main.Select n peaks from A i.e. P = ψ(A), where
P are n peaks extracted from the A and ψ function
which finds n global maxima from A
3. Peak Filtering: We filter peaks at two levels:
• We filter all n peaks P in the transform domain
to obtain intermediate peaks pairs IP that sat-
isfy certain geometrical constraints. i.e. IP =
C
A
(P), where C
A
represents the parallel and or-
thogonal constraints in the transform space.
• We further filter all intermediate peaks pairs IP,
in the form of line segments in the spatial co-
ordinate domain to get the selected peak pairs
PP, i.e. PP = C
I
(IP) that form potential rect-
angles. where C
I
are spatial constraints in the
image domain.
Having transformed the original edge image E us-
ing one of the transformations mentioned above, we
have A. Our next step is to identify n peaks in the
transform space that correspond to lines in the image.
Since A is a type of accumulator, the simplest mech-
anism of extracting peaks is using thresholding. That
is, A > t
p
, where t
p
is a pre-defined threshold value,
which if needed can be learnt from training.
It is possible that there can be more than one peak
in the transform space that all satisfy the threhsol-
ding constraint within any reasonable neighborhood.
These multiple peaks of the transform space within
a small neighborhood region will correspond to du-
plicate lines in the spatial domain. Duplicated peaks
will also be discarded in the Transform Domain Peak
Filtering stage. We therefore, discard all other peaks
within any specified neighborhood region.
We extend the algorithm of (Jung and Schramm,
2004), to detect rectangle patterns in the trans-
form space. We denote the n extracted peaks as
P
1
, P
2
, ... P
n
of A(ρ,θ). We iteratively compare peaks
P
i
and P
j
in order to identify those that are parallel
to one another, i.e. those that satisfy the following
conditions:
4θ = |θ
i
− θ
j
| < t
θ
|A(ρ
i
, θ
i
) − A(ρ
j
, θ
j
)| < t
l
A(ρ
i
, θ
i
) + A(ρ
j
, θ
j
)
2
(2)
where, t
θ
is an angular threshold that determines if
P
i
and P
j
correspond to parallel lines. t
l
is the thresh-
old that determines if lines corresponding to P
i
and
P
j
are of the same lengths. Each pair of peaks that
satisfy 2 denote extended peaks EP = (β
k
, A
k
), where
β
k
=
1
2
(θ
i
+ θ
j
) and A
k
=
1
2
(A(ρ
i
, θ
i
) + A(ρ
j
, θ
j
)) (3)
These peaks are further checked for orthogonality.
The lines that correspond to these extended peaks are
orthogonal if the following condition is satisfied:
4β = ||β
r
− β
s
| − 90
o
| < t
β
(4)
We call those extended peak pairs that satisfy con-
dition 4 as the intermediate peak pairs IP. These IP’s,
representing potential candidate rectangles, that are
further filtered in the spatial domain, in order to ex-
tract the actual ones.
Let us consider one intermediate peak pair IP
i
(con-
taining 4 peaks, pair of 2 parallel peaks that are or-
thogonal to one another) and the corresponding line
segments IL
i
in the spatial domain. We use a line
VISAPP 2011 - International Conference on Computer Vision Theory and Applications
216
intersection mechanism for detecting genuine rectan-
gular structures subtended by these line segments IL
i
.
According to our technique,
• We compute the point of intersection for each
pair of orthogonal line segments from the set IL.
• For each intersection point , we check if:
|
−−−→
IL
s
i
.
−−−→
IL
f
i
||
−−→
IL
s
i
.
−−→
IL
f
i
||
2
| < ξ (5)
where, IL
s
i
and IL
f
i
are the end points of line seg-
ment IL
i
.
intersection mechanism for detecting genuine rectan-
gular structures subtended by these line segments IL
i
.
According to our technique,
• We compute the point of intersection for each
pair of orthogonal line segments from the set IL.
• For each intersection point , we check if:
|
−−→
IL
s
i
.
−−→
IL
f
i
||
−→
IL
s
i
.
−→
IL
f
i
||
2
| < ξ (5)
where, IL
s
i
and IL
f
i
are the end points of line seg-
ment IL
i
.
a b
Figure 3: Legal (a) and False (b) rectangular structures
formed from intersection of intermediate lines
The satisfaction of Equation 5 guarantees that the
point of intersection lies in between the end points of
the corresponding line segments. This constraint al-
lows us distinguish between genuine rectangular pat-
ters as in Figure 3a from incorrect ones as in Fig-
ure 3b. The choice of the value of our threshold ξ in-
troduces additional relaxations on the constraint. We
have chosen the value of ξ to be 10 in all our ex-
periments. Within this framework, it is also possible
introduce application based size constraints on rect-
angles being detected. This technique of spatial fil-
tering of peaks for detecting rectangles allow robust
detections of contiguous blocks of rectangles simul-
taneously.
2.1 SIFT Descriptors & Classification
Scale Invariant Feature Transform (SIFT) is an
approach for detecting and extracting local feature
descriptors that are reasonably invariant to changes in
illumination, image noise, rotation, scaling, and small
changes in viewpoint. In our framework, we use
the SIFT interest point detector proposed by (Lowe,
2004). From every parking space extracted using
the rectangular parking space detection algorithm
mentioned above, we extract key points and their
corresponding SIFT descriptors. We acknowledge
that other descriptors can be used, however, the
performance of the system will depend on the robust-
ness of the chosen descriptor. The SIFT descriptor
is a 128-dimensional vector containing a set of
gradient orientation histograms for every key point
extracted for that image. We notice that for images
containing occupied parking spaces, the number of
set of key points that can be extracted is much larger
than the number of key points that could be located
on an empty parking space. We will exploit this
characteristic during the classification. We illustrate
the locations of some of these extracted descriptors
in images as shown in Figure4.
Figure 4: Locations of SIFT descriptors extracted from the
rectangular parking spaces
For the purposes of image classification, using
the high dimensional SIFT descriptors can often by
an expensive procedure. Therefore, it is common to
perform dimensionality reduction so that higher or-
der features can be computed from these SIFT de-
scriptors. Here, we cluster a large training set of
descriptors sampled from our data set using the k-
means clustering algorithm and quantize these orig-
inal 128-dimensional SIFT descriptors by assigning a
label of the closest center as described in the work of
(Yang and Newsam, 2008). As in (Yang and Newsam,
2008), we associate the frequency count of these la-
bels to each higher order features. In our work, we
use a set of 50 higher order features extracted from
the SIFT descriptors.
In the final step, we perform classification of the SIFT
based features using supervised classification algo-
rithms. We classify the detected parking spaces and
their corresponding SIFT features using the Support
Vector Machine (SVM) classifier. In addition, we
also compare the results of the SVM classifier to a
simple thresholding operator for empty parking space
detection. SVM is a supervised learning algorithm
that determines a hyperplane that separates classes
by maximizing the margins between them (Yang and
Newsam, 2008). SVM is easy to implement and use
and have been proved to be very useful in handling
high-dimensional feature vectors as in our case. In
order to train our SVM, we use manually categorized,
sizeable collection of positive (empty parking spaces)
and negative (occupied parking spaces) examples. We
apply a 3-fold cross validation to estimate the optimal
parameters.
In comparison to the SVM type supervised classifi-
cation, we also implement a simple thresholding op-
erator that categorizes parking spaces as empty if the
number of keypoints extracted from the SIFT analy-
sis is less than a pre-defined threshold. We perform
Figure 3: Legal (a) and False (b) rectangular structures
formed from intersection of intermediate lines.
The satisfaction of Equation 5 guarantees that the
point of intersection lies in between the end points of
the corresponding line segments. This constraint al-
lows us distinguish between genuine rectangular pat-
ters as in Figure 3a from incorrect ones as in Fig-
ure 3b. The choice of the value of our threshold ξ in-
troduces additional relaxations on the constraint. We
have chosen the value of ξ to be 10 in all our ex-
periments. Within this framework, it is also possible
introduce application based size constraints on rect-
angles being detected. This technique of spatial fil-
tering of peaks for detecting rectangles allow robust
detections of contiguous blocks of rectangles simul-
taneously.
2.1 SIFT Descriptors & Classification
Scale Invariant Feature Transform (SIFT) is an ap-
proach for detecting and extracting local feature de-
scriptors that are reasonably invariant to changes in
illumination, image noise, rotation, scaling, and small
changes in viewpoint. In our framework, we use
the SIFT interest point detector proposed by (Lowe,
2004). From every parking space extracted using the
rectangular parking space detection algorithm men-
tioned above, we extract key points and their cor-
responding SIFT descriptors. We acknowledge that
other descriptors can be used, however, the perfor-
mance of the system will depend on the robustness
of the chosen descriptor. The SIFT descriptor is a
128-dimensional vector containing a set of gradient
orientation histograms for every key point extracted
for that image. We notice that for images contain-
ing occupied parking spaces, the number of set of
key points that can be extracted is much larger than
the number of key points that could be located on an
empty parking space. We will exploit this character-
istic during the classification. We illustrate the loca-
tions of some of these extracted descriptors in images
as shown in Figure4.
intersection mechanism for detecting genuine rectan-
gular structures subtended by these line segments IL
i
.
According to our technique,
• We compute the point of intersection for each
pair of orthogonal line segments from the set IL.
• For each intersection point , we check if:
|
−−→
IL
s
i
.
−−→
IL
f
i
||
−→
IL
s
i
.
−→
IL
f
i
||
2
| < ξ (5)
where, IL
s
i
and IL
f
i
are the end points of line seg-
ment IL
i
.
a b
Figure 3: Legal (a) and False (b) rectangular structures
formed from intersection of intermediate lines
The satisfaction of Equation 5 guarantees that the
point of intersection lies in between the end points of
the corresponding line segments. This constraint al-
lows us distinguish between genuine rectangular pat-
ters as in Figure 3a from incorrect ones as in Fig-
ure 3b. The choice of the value of our threshold ξ in-
troduces additional relaxations on the constraint. We
have chosen the value of ξ to be 10 in all our ex-
periments. Within this framework, it is also possible
introduce application based size constraints on rect-
angles being detected. This technique of spatial fil-
tering of peaks for detecting rectangles allow robust
detections of contiguous blocks of rectangles simul-
taneously.
2.1 SIFT Descriptors & Classification
Scale Invariant Feature Transform (SIFT) is an
approach for detecting and extracting local feature
descriptors that are reasonably invariant to changes in
illumination, image noise, rotation, scaling, and small
changes in viewpoint. In our framework, we use
the SIFT interest point detector proposed by (Lowe,
2004). From every parking space extracted using
the rectangular parking space detection algorithm
mentioned above, we extract key points and their
corresponding SIFT descriptors. We acknowledge
that other descriptors can be used, however, the
performance of the system will depend on the robust-
ness of the chosen descriptor. The SIFT descriptor
is a 128-dimensional vector containing a set of
gradient orientation histograms for every key point
extracted for that image. We notice that for images
containing occupied parking spaces, the number of
set of key points that can be extracted is much larger
than the number of key points that could be located
on an empty parking space. We will exploit this
characteristic during the classification. We illustrate
the locations of some of these extracted descriptors
in images as shown in Figure4.
Figure 4: Locations of SIFT descriptors extracted from the
rectangular parking spaces
For the purposes of image classification, using
the high dimensional SIFT descriptors can often by
an expensive procedure. Therefore, it is common to
perform dimensionality reduction so that higher or-
der features can be computed from these SIFT de-
scriptors. Here, we cluster a large training set of
descriptors sampled from our data set using the k-
means clustering algorithm and quantize these orig-
inal 128-dimensional SIFT descriptors by assigning a
label of the closest center as described in the work of
(Yang and Newsam, 2008). As in (Yang and Newsam,
2008), we associate the frequency count of these la-
bels to each higher order features. In our work, we
use a set of 50 higher order features extracted from
the SIFT descriptors.
In the final step, we perform classification of the SIFT
based features using supervised classification algo-
rithms. We classify the detected parking spaces and
their corresponding SIFT features using the Support
Vector Machine (SVM) classifier. In addition, we
also compare the results of the SVM classifier to a
simple thresholding operator for empty parking space
detection. SVM is a supervised learning algorithm
that determines a hyperplane that separates classes
by maximizing the margins between them (Yang and
Newsam, 2008). SVM is easy to implement and use
and have been proved to be very useful in handling
high-dimensional feature vectors as in our case. In
order to train our SVM, we use manually categorized,
sizeable collection of positive (empty parking spaces)
and negative (occupied parking spaces) examples. We
apply a 3-fold cross validation to estimate the optimal
parameters.
In comparison to the SVM type supervised classifi-
cation, we also implement a simple thresholding op-
erator that categorizes parking spaces as empty if the
number of keypoints extracted from the SIFT analy-
sis is less than a pre-defined threshold. We perform
Figure 4: Locations of SIFT descriptors extracted from the
rectangular parking spaces.
For the purposes of image classification, using
the high dimensional SIFT descriptors can often by
an expensive procedure. Therefore, it is common to
perform dimensionality reduction so that higher or-
der features can be computed from these SIFT de-
scriptors. Here, we cluster a large training set of
descriptors sampled from our data set using the k-
means clustering algorithm and quantize these orig-
inal 128-dimensional SIFT descriptors by assigning a
label of the closest center as described in the work of
(Yang and Newsam, 2008). As in (Yang and Newsam,
2008), we associate the frequency count of these la-
bels to each higher order features. In our work, we
use a set of 50 higher order features extracted from
the SIFT descriptors.
In the final step, we perform classification of the
SIFT based features using supervised classification
algorithms. We classify the detected parking spaces
and their corresponding SIFT features using the Sup-
port Vector Machine (SVM) classifier. In addition,
we also compare the results of the SVM classifier to a
simple thresholding operator for empty parking space
detection. SVM is a supervised learning algorithm
that determines a hyperplane that separates classes
by maximizing the margins between them (Yang and
Newsam, 2008). SVM is easy to implement and use
and have been proved to be very useful in handling
high-dimensional feature vectors as in our case. In
order to train our SVM, we use manually categorized,
sizeable collection of positive (empty parking spaces)
and negative (occupied parking spaces) examples. We
apply a 3-fold cross validation to estimate the optimal
parameters.
RECTANGULAR EMPTY PARKING SPACE DETECTION USING SIFT BASED CLASSIFICATION
217
In comparison to the SVM type supervised classi-
fication, we also implement a simple thresholding op-
erator that categorizes parking spaces as empty if the
number of keypoints extracted from the SIFT analy-
sis is less than a pre-defined threshold. We perform
a number of initial experimentations to determine an
optimal threshold for this binary classification task.
3 EXPERIMENTATION
& RESULTS
Dataset Description: A large collection of both
synthetic and real data is used for evaluating our
model. In the synthetic set, we use a collection of
60 images at varying resolutions starting from 256 x
256 pixels to 493 x 403 pixels. For real images, we
again use a collection of 60 images captured aerially
from the top of a building using a CANON G7 10
Mega pixel camera. We manually ground truth all
the data in the synthetic set containing a total of
272 occupied parking spaces and 136 empty parking
spaces. Similarly, we ground truth the real data set
to contain a total of 285 occupied parking spaces and
161 empty parking spaces.
Performance Metrics: To evaluate the performance
of classification, we compute the following perfor-
mance metrics using a two-way contingency table as
in (Seo and Urmson, 2009).
• Precision, p =
a
a+b
, if a + b > 0, otherwise unde-
fined
• Recall, r =
a
a+c
, if a + c > 0, otherwise undefined
• False Positives, f p =
b
b+d
, if b + d > 0, otherwise
undefined
• False Negatives, f n =
c
a+c
, if a + c > 0, otherwise
undefined
• Accuracy, acc =
a+d
a+b+c+d
, if a + d > 0, otherwise
undefined
We begin by dividing the synthetic and real data
into 3 sets containing 20 images each for training,
validation and testing. During the training process,
we manually extract rectangular parking spaces,
perform SIFT analysis thus computing the 128
dimension local SIFT descriptors. Furthermore, we
extract 50 dimensional higher order features from the
128 dimension descriptors using the k-means clus-
tering approach described in section 2.1. Finally, we
associate these features to the respective ground-truth
class label and thus train the SVM classifier. We
evaluate the feature and classifier combination using
Table 1: Compared mean (3 fold) metric values between
SVM and Thresholding classification on synthetic dataset.
Metric SVM Thresholding
Precision 0.969 0.99
Recall 0.942 0.936
False Positives 0.032 0.020
False Negatives 0.031 0.021
Accuracy 0.969 0.975
Table 2: Compared mean (3 fold) metric values between
SVM and Thresholding classification on real dataset.
Metric SVM Thresholding
Precision 0.952 0.965
Recall 0.932 0.947
False Positive 0.093 0.063
False Negatives 0.061 0.043
Accuracy 0.929 0.969
a manually intensive 3 fold cross validation.
Synthetic Data: We perform experiments on syn-
thetic data containing clear rectangular parking
spaces created under controlled conditions. The im-
ages are free from noise variations and will simulate
the effect of various parking scenarios. The images
in the test sets are taken through steps of edge detec-
tion using a Canny operator with threshold 0.1 fol-
lowed by edge enhancement. We then perform RPSD
to detect and extract rectangular parking spaces. The
mean error in detecting rectangular parking spaces us-
ing the proposed algorithm on the synthetic dataset
was found to be around 4.85%.
The detected parking spaces are then subjected
to SIFT analysis and classified using both the SVM
and Thresholding approaches. We tabulate the
quantitative evaluation of the performance of the
classifiers in Table 1.
Real Data: We test the same proposed model on the
set of real data. The images in the real data are more
challenging with variations due to noise, varying il-
lumination conditions etc. Table 2 shows the accu-
racy of the classifiers. We would like to highlight
that the proportion of missed detections of the RPSD
model on real images was found to be 12%. This is
an increase of 7.15% in comparison to the detections
on the synthetic data. However, the proposed model
maintains an accuracy greater than 95%.
We notice from the results on both synthetic and
real data that the simple threshold based classifier can
produce better classification accuracy than the SVM
classifier. However the threshold value is fixed on
VISAPP 2011 - International Conference on Computer Vision Theory and Applications
218
a trial and error basis. On the other hand, the SVM
classifier is more systematic approach that gives
additional flexibility of easily extending the current
binary classification problem into a multi-class prob-
lem. One of the other main problems of the proposed
framework is the proportion of missed detections
of the RPSD model. The following experiment
examines the impact of the RPSD algorithm on the
accuracy of the model. We summarize the results of
the RPSD algorithm on some example synthetic and
real-time images in Figure 5.
Figure 5: Results of rectangular parking space detection on
synthetic and real images.
We also plot the precision-recall curves of model both
on synthetic and real data. This is illustrated in Fig-
ure 6. We have computed the break-even-point to be
at 0.947 for the real data and 0.963 for the synthetic
data,which is indicative of a robust and accurate sys-
tem.
Figure 5: Results of rectangular parking space detection on
synthetic and real images
be at 0.947 for the real data and 0.963 for the synthetic
data,which is indicative of a robust and accurate sys-
tem.
Figure 6: Precision versus Recall curves for Real and Syn-
thetic Data
We attribute the reduced detection rate of the
rectangular parking space detection algorithm to two
main reasons. The RPSD algorithm depends on the
parking lines to detect parking spaces and therefore
partial or full occlusion of these parking lines can
affect the detection process. In addition, the RPSD
algorithm is aimed at detecting rectangular regions
alone. However, in most real images, the parking re-
gions are not all rectangular. This limitation is fairly
easy to address by incorporating changes into the con-
dition in Equation 2, such that small variations in the
angular orientations of the parking lines is made ac-
ceptable.
4 Conclusions
We have proposed a framework for detecting
empty parking spaces in images. Our method com-
bines Radon transform based rectangular parking
space detection scheme with SIFT analysis for clas-
sification of empty parking spaces. The results of ap-
plying our technique to synthetic and real data sug-
gests that the proposed framework is more accurate
and robust to the presence of noise, clutter and illu-
mination changes in images. In our future work we
plan to extend the proposed technique to detect park-
ing spaces of other arbitrary shapes and compare them
with other state-of-the-art methods.
REFERENCES
Benson, J. P., Donovan, T. O., Sullivan, P. O., Roedig, U.,
and Sreenan, C. (2006). Car-park management using
wireless sensor networks. Proc. 31st IEEE Conf. Lo-
cal Computer Networks, pages 588–595.
Bhaskar, H., Werghi, N., and Mansoori, S. A. (2010). Com-
bined spatial and transform domain analysis for rect-
angle detection. Proc. of the 13th IET Conference on
Information FUSION.
Bong, D. B. L., Ting, K. C., and Lai, K. C. (2008). In-
tegrated approach in the design of car park occu-
pancy information system (coins). IAENG Interna-
tional Journal of Computer Science, 1(1):7–14.
Funck, S., Mohler, N., and Oertel, W. (2004). Determining
car-park occupancy from single images. IEEE Intelli-
gent Vehicles Symposium, pages 325–328.
Jung, C. R. and Schramm, R. (2004). Rectangle detection
based on a windowed hough transform. SIBGRAPI
’04: Proc. of the Computer Graphics and Image Pro-
cessing, XVII Brazilian Symposium, pages 113–120.
Lowe, D. (2004). Distinctive image features from scalein-
variant keypoints. International Journal of Computer
Vision, 60(2):91–110.
Ristola, T. (1992). Parking guidance system in tapiola.
Proc. IEE Conf. Road Traffic Monitoring, page 195.
Seo, Y.-W., Ratliff, N., and Urmson, C. (2009). Self-
supervised aerial image analysis for extracting park-
ing lot structure. Proc. of International Joint Confer-
ence on Artificial Intelligence.
Seo, Y.-W. and Urmson, C. (2009). A hierarchical im-
age analysis for extracting parking lot structures from
aerial images. Robotics Institute, Carnegie Mellon
University: CMU-RI-TR-09-03.
Tang, V. W. S., Zheng, Y., and Cao, J. (2006). An intelligent
car park management system based on wireless sensor
networks. Proc. Int. Sym. Pervasive Computing and
Applications, pages 65–70.
Wang, X. and Hanson, A. (1998). Parking lot analysis
and visualization from aerial images. Proc. of the 4th
IEEE Workshop on Applications of Computer Vision,
page 36.
Wolff, J., Heuer, T., Gao, H., Weinmann, M., Voit, S., and
Hartmann, U. (2006). Parking monitor system based
on magnetic field sensors. Proc. IEEE Conf. Intelli-
gent Transportation Systems, pages 1275–1279.
Yang, Y. and Newsam, S. (2008). Comparing sift descrip-
tors and gabor texture features for classification of re-
mote sensed imagery. Proc. IEEE International Con-
ference on Image Processing, pages 1852–1855.
Yu, C. and Liu, J. (2004). A type of sensor to detect occu-
pancy of vehicle berth in carpark. Proc. 7th Int. Conf.
Signal Processing, pages 2708–2711.
Figure 6: Precision versus Recall curves for Real and Syn-
thetic Data.
We attribute the reduced detection rate of the rectan-
gular parking space detection algorithm to two main
reasons. The RPSD algorithm depends on the parking
lines to detect parking spaces and therefore partial or
full occlusion of these parking lines can affect the de-
tection process. In addition, the RPSD algorithm is
aimed at detecting rectangular regions alone. How-
ever, in most real images, the parking regions are not
all rectangular. This limitation is fairly easy to ad-
dress by incorporating changes into the condition in
Equation 2, such that small variations in the angular
orientations of the parking lines is made acceptable.
4 CONCLUSIONS
We have proposed a framework for detecting empty
parking spaces in images. Our method combines
Radon transform based rectangular parking space de-
tection scheme with SIFT analysis for classification
of empty parking spaces. The results of applying
our technique to synthetic and real data suggests that
the proposed framework is more accurate and ro-
bust to the presence of noise, clutter and illumination
changes in images. In our future work we plan to ex-
tend the proposed technique to detect parking spaces
of other arbitrary shapes and compare them with other
state-of-the-art methods.
REFERENCES
Benson, J. P., Donovan, T. O., Sullivan, P. O., Roedig, U.,
and Sreenan, C. (2006). Car-park management using
wireless sensor networks. Proc. 31st IEEE Conf. Lo-
cal Computer Networks, pages 588595.
Bhaskar, H., Werghi, N., and Mansoori, S. A. (2010). Com-
bined spatial and transform domain analysis for rect-
angle detection. Proc. of the 13th IET Conference on
Information FUSION.
Bong, D. B. L., Ting, K. C., and Lai, K. C. (2008). In-
tegrated approach in the design of car park occu-
pancy information system(coins). IAENG Interna-
tional Journal of Computer Science, 1(1):714.
Funck, S., Mohler, N., and Oertel, W. (2004). Determining
car-park occupancy from single images. IEEE Intelli-
gent Vehicles Symposium, pages 325328.
Jung, C. R. and Schramm, R. (2004). Rectangle detection
based on a windowed hough transform. SIBGRAPI
04: Proc. of the Computer Graphics and Image Pro-
cessing, XVII Brazilian Symposium, pages 113120.
Lowe, D. (2004). Distinctive image features from scalein-
variant keypoints. International Journal of Computer
Vision, 60(2):91110.
Ristola, T. (1992). Parking guidance system in tapiola.
Proc. IEE Conf. Road Traffic Monitoring, page 195.
Seo, Y.-W., Ratliff, N., and Urmson, C. (2009). Self-
supervised aerial image analysis for extracting park-
ing lot structure. Proc. of International Joint Confer-
ence on Artificial Intelligence.
Seo, Y.-W. and Urmson, C. (2009). A hierarchical im-
age analysis for extracting parking lot structures from
aerial images. Robotics Institute, Carnegie Mellon
University: CMU-RI-TR-09-03.
RECTANGULAR EMPTY PARKING SPACE DETECTION USING SIFT BASED CLASSIFICATION
219
Tang, V. W. S., Zheng, Y., and Cao, J. (2006). An intelligent
car park management system based on wireless sensor
networks. Proc. Int. Sym. Pervasive Computing and
Applications, pages 6570.
Wang, X. and Hanson, A. (1998). Parking lot analysis
and visualization from aerial images. Proc. of the 4th
IEEE Workshop on Applications of Computer Vision,
page 36.
Wolff, J., Heuer, T., Gao, H., Weinmann, M., Voit, S., and
Hartmann, U. (2006). Parking monitor system based
on magnetic field sensors. Proc. IEEE Conf. Intelli-
gent Transportation Systems, pages 12751279.
Yang, Y. and Newsam, S. (2008). Comparing sift descrip-
tors and gabor texture features for classification of re-
mote sensed imagery. Proc.IEEE International Con-
ference on Image Processing, pages 18521855.
Yu, C. and Liu, J. (2004). A type of sensor to detect occu-
pancy of vehicle berth in carpark. Proc. 7th Int. Conf.
Signal Processing, pages 27082711.
VISAPP 2011 - International Conference on Computer Vision Theory and Applications
220