IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 18, NO.

6, JUNE 2017 1393

A Video-Based System for Vehicle Speed

Measurement in Urban Roadways
Diogo Carbonera Luvizon, Bogdan Tomoyuki Nassu, and Rodrigo Minetto

Abstract—In this paper, we propose a nonintrusive video-based

system for vehicle speed measurement in urban roadways. Our
system uses an optimized motion detector and a novel text de-
tector to efficiently locate vehicle license plates in image regions
containing motion. Distinctive features are then selected on the
license plate regions, tracked across multiple frames, and rectified
for perspective distortion. Vehicle speed is measured by comparing
the trajectories of the tracked features to known real-world mea-
sures. The proposed system was tested on a data set containing
approximately 5 h of videos recorded in different weather condi-
tions by a single low-cost camera, with associated ground truth
speeds obtained by an inductive loop detector. Our data set is
freely available for research purposes. The measured speeds have
an average error of −0.5 km/h, staying inside the [−3, +2] km/h
limit determined by regulatory authorities in several countries in
over 96.0% of the cases. To the authors’ knowledge, there are Fig. 1. System setup.
no other video-based systems able to achieve results comparable
to those produced by an inductive loop detector. We also show
that our license plate detector outperforms two other published
state-of-the-art text detectors, as well as a well-known license plate
detector, achieving a precision of 0.93 and a recall of 0.87.
Index Terms—vehicle speed measurement, license plate detec-
tion, feature tracking, vehicle motion detection.

I. I NTRODUCTION

S YSTEMS for vehicle detection and speed measurement

play an important role in enforcing speed limits. They
also provide relevant data for traffic control. Those systems
are divided in intrusive and non-intrusive [1]. Intrusive sensors,
usually based on inductive loop detectors, are widely used,
but have complex installation and maintenance, accelerate as- Fig. 2. Sample image captured by our system.
phalt deterioration, and can be damaged by wear and tear.
Non-intrusive sensors, which include laser meters and Doppler connected to video cameras [2] that record the license plates of
radars, avoid these problems, but are usually more expensive vehicles that exceed the speed limit — thus, the infrastructure
and require frequent maintenance. As digital cameras become for such systems is already available in most cases.
cheaper and able to produce images with higher quality, video- In this paper we describe the pipeline for a non-intrusive
based systems can become a lower cost alternative for non- video-based system for vehicle speed measurement in urban
intrusive speed measurement. In fact, existing systems are often roadways. Our goal is measuring vehicle speeds with accuracy
comparable to that obtained by a system based on inductive
loop detectors. The input video is captured by a single fixed
Manuscript received February 12, 2016; revised May 26, 2016; accepted
August 29, 2016. Date of publication September 26, 2016; date of current overhead camera, positioned so that the rear license plate of
version May 29, 2017. This work was supported in part by the Brazilian Agency vehicles in three adjacent lanes are clearly visible, as shown
Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) in Fig. 1. A sample image from this setup is shown in Fig. 2.
under Grant 444789/2014-6. The Associate Editor for this paper was J. Zhang.
D. C. Luvizon was with the Department of Informatics, Federal University This setup allows the same images to be used for both speed
of Technology of Paraná (UTFPR), 80230-901 Curitiba-PR, Brazil. He is now measurement and license plate identification (e.g. for locating
with Université de Cergy-Pontoise, 95000 Cergy-Pontoise, France (e-mail: stolen vehicles, or in the case of a speed limit violation).
[email protected]).
B. T. Nassu and R. Minetto are with the Department of Informatics, Federal We make some assumptions about the scene and the problem
University of Technology of Paraná (UTFPR), 80230-901 Curitiba-PR, Brazil domain: video frames are equally spaced in time; each lane
(e-mail: [email protected]; [email protected]). lies on a plane; the vehicles move at a constant speed and
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. with a straight trajectory from the lower to the upper part
Digital Object Identifier 10.1109/TITS.2016.2606369 of the image; and the license plates are at approximately the
1524-9050 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1394 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 18, NO. 6, JUNE 2017

measurement methods, respectively. Experimental evaluation

and results are reported in Section VII. Finally, in Section VIII
we state the conclusions.

II. R ELATED W ORK

A. Vehicle Speed Estimation and Measurement
Several video-based approaches were proposed for estimat-
ing or measuring the speed of vehicles in roadways. Most
methods include a background/foreground segmentation step to
Fig. 3. Overview of the proposed system. detect image regions containing motion. Common approaches
for this task include simple frame differences [7]–[10], as
same distance from the ground. These assumptions allow us well as statistic models based on medians [11]–[13], Gaussian
to measure vehicle speeds without modeling the 3-D space, or distributions [14] or other measures [15]. Vehicle speeds are
requiring precise camera calibration or positioning. estimated by tracking image features or regions, including blobs
The proposed system works by tracking sets of distinctive [12], image patches [14], edges [7], [11], corners [8], [9], the
features extracted from image regions around each vehicle’s license plate region [16]–[18], or a combination of such features
license plate, and is divided into five main parts, as shown [13]. Rectification for perspective distortion is also a step found
in Fig. 3. Initially, an optimized motion detection algorithm in most methods, and may occur before or after feature tracking.
identifies image regions containing moving vehicles. These Although the cited methods have some steps in common,
regions are fed to a novel license plate detector, which re- they also have fundamental differences, not only on the way they
turns a set of axis-aligned rectangular sub-images around the measure vehicle speeds, but also on the type of scenario they
vehicles’ license plates. Features are then extracted from each can be used.
sub-image [3], and tracked using the Kanade–Lucas–Tomasi Methods based on direct blob analysis [7], [8], [10], [12],
(KLT) algorithm [4]. To cope with large displacements, from [14], [15] are sensitive to conditions such as shadows, per-
vehicles moving at high speeds, an initial motion estimation spective, and illumination variations. Moreover, these methods
is performed by matching features extracted by the Scale- produce satisfactory results only when the camera is positioned
Invariant Feature Transform [5]. Finally, vehicle speeds are high above the roadway, with the blobs being tracked for many
measured by comparing the trajectories of the tracked features frames. The same issues affect methods which use other types
to known real world measures. of features, but still compute them from blobs, such as those
A proof-of-concept of our system was evaluated on approx- proposed by Zhiwei et al. [11], which detects edges near the
imately five hours of videos in different weather and recording limits of each blob, or Palaio et al. [13], which extracts from
conditions. The videos have an associated ground truth dataset each blob features such as derivatives, Laplacian and color. As
containing vehicle speeds measured by a high precision system discussed in Section VII, we have compared our system with
based on inductive loop detectors, properly calibrated and ap- a blob tracking approach based on a particle filter, similar in
proved by the Brazilian national metrology agency (Inmetro). concept to the one proposed by Maduro et al. [12].
This data set is itself a contribution of our work, and can be The method from Dogan et al. [9] avoids the problems
freely obtained for research purposes.1 Our system was able associated with blob analysis by directly tracking distinctive
to measure speeds with an average error of −0.5 km/h, stay- features using the Lucas–Kanade optical flow algorithm [4].
ing inside the [−3, +2] km/h limit determined by regulatory However, their method assumes that all the tracked features
authorities in several countries, in over 96.0% of the cases. belong to the same vehicle — thus it can handle only a single
We also show that our license plate detector outperforms other vehicle at a time. Moreover, they do not take perspective into
two published state-of-the-art text detectors, as well as a well- account, and require a side view of the vehicles.
known license plate detector. The work from Garibotto et al. [16] relies on how characters
A preliminary version of the system described here was pub- detected by an optical character recognition (OCR) algorithm
lished at the 2014 IEEE International Conference on Acoustics, vary in size and position. Their method demands a very robust
Speech and Signal Processing (ICASSP) [6]. The system de- OCR, and did not produce acceptable results even in a con-
scribed in the present paper differs from that version in several trolled environment — average errors for a single-camera setup
aspects, such as the inclusion of an initial motion detection ranged from 3% to 13%. A similar issue was observed in the
step, a new algorithm for license plate detection, and refined work from Czajewski and Iwanowski [17], which is also based
computations for the final measured speed. The new system was on license plate recognition. Note that, although our system has
evaluated on a broader and deeper manner. a license plate detection step, it does not require the characters
The rest of this paper is divided as follows. In Section II, we to be precisely segmented or recognized.
discuss related work. In Sections III– VI we present our motion
detection, license plate detection, vehicle tracking and speed B. License Plate Detection
1 The full dataset will be made available at the time of publication, a sample Although license plate detection is not our primary concern
is available at www.dainf.ct.utfpr.edu.br/%7erminetto/projects/vehicle-speed. in this work, it is a necessary step for the proposed speed
LUVIZON et al.: VIDEO-BASED SYSTEM FOR VEHICLE SPEED MEASUREMENT IN URBAN ROADWAYS 1395

Fig. 5. Routine to find hills in Ψ.

Fig. 4. Motion detection: regions of interest are delimited based on a fore-

ground image mask and a vertical projection profile. where D are the binary images obtained from thresholded frame
differences, and the τ parameter represents the duration of the
measurement system. The surveys from Anagnostopoulos et al. expected motion in frame units. In our tests, we used τ = 5.
[19] and Du et al. [20] review state-of-the-art license The binary segmentation mask M is obtained by
plate detection algorithms up to 2013. Those algorithms 
rely on attributes such as edges, texture, color, shape, and 1 if H(x, y, t) > 0
M(x, y, t) = (2)
geometry—extracted using techniques such as the Sobel op- 0 otherwise.
erator [18], [21], the Canny detector [22], conditional random
field [23], template-matching [24], wavelets [25], or the Scale An example is shown in Fig. 4. Note that we do not have to
Invariant Feature Transform (SIFT) [26]. The faced challenges identify vehicle boundaries in a precise manner, as our system
include poor maintenance, occlusion, variations in position and does not rely on blob analysis.
illumination, complex backgrounds, low contrast, low image To reduce the processing time when generating image M
resolution, and motion blur. and in subsequent steps, the images are subsampled — i.e. the
As detailed in Section VII, the specialized license plate values of x and y in (1) are restricted to a regular sparse grid,
detector employed by our system was compared with three with pixels outside the grid being skipped. The segmentation
other alternatives. The Zheng et al. algorithm [21] uses a may become less precise, but this is acceptable as long as the
rectangular sliding window to identify image regions with high vehicle license plate remains entirely within its corresponding
gradient density, which probably contain license plates. The region. In tests involving the complete system, we observed that
Stroke Width Transform (SWT), by Epshtein et al. [22], is a a subsampling factor of 4 in both axes (i.e. processing 1 of each
text detector based on the orientations of the gradients over 16 pixels) reduced the processing time to 16.47% of the original
edge pixels, which are used to determine a local stroke width time, without any loss in detection performance.
for candidate characters. SnooperText, by Minetto et al. [27], After the sub-sampled binary segmentation mask M is ob-
is a multi-scale text detector that uses morphological image tained, we perform a vertical projection profile analysis [30]
segmentation and character/non-character classification based to separate vehicles horizontally. We take the lower part of
on shape descriptors. SnooperText validates candidate text re- image M, which shows the region closer to the camera, and
gions using the T-HOG classifier [28], a specialized gradient- count the foreground pixels in each column, generating a his-
based descriptor tuned for single-line text regions. All these togram with n bins (for n image columns). This histogram is
algorithms include steps of pre-processing and tests for filtering smoothed, to reduce noise, and interpreted as an array Ψ.
out incorrect results, based on size and geometry. Fig. 4 shows an example of vertical projection profile. It can
be seen that the interval containing a vehicle is delimited by an
III. M OTION D ETECTION ascending and a descending slope, corresponding respectively
The first step in our system’s pipeline is detecting moving to the left and right boundaries of the vehicle. The F IND -H ILLS
vehicles, limiting further processing to a set of regions of routine (Fig. 5) is used to determine these boundaries for each
interest. Ideally, each region of interest must contain the entire vehicle. It receives as parameters the projection profile array Ψ,
license plate from a single vehicle. An overview of the motion and a threshold ρ (0.1, in our tests) that defines the minimum
detection approach that we develop is shown in Fig. 4. angle of inclination for a boundary. It returns a pair of lists
Motion detection begins with a rough foreground/ {A, D}, such that each list element represents a hill’s ascending
background segmentation. We use the Motion History Image and descending border, respectively.
(MHI) concept from Bobick and Davis [29]. The MHI H for In step 2 of F IND -H ILLS, we call the F IND -I NCLINATION
time t is given by routine, outlined as Fig. 6. The purpose of this routine is to
 determine the rising and falling phases of Ψ [see Fig. 8(a)],
τ if D(x, y, t) = 1 given as arrays R and F , respectively. The threshold 0 ≤ ρ ≤ 1
H(x, y, t) = (1)
max (0, H(x, y, t − 1) − 1) otherwise. is used to discard false phases: since vehicle regions are
1396 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 18, NO. 6, JUNE 2017

Fig. 6. Routine to find rising and falling phases in Ψ.

Fig. 8. Internal steps of the F IND -H ILLS routine. (a) Rising and falling
phases according to a given threshold ρ. (b) Ascending and descending slopes.
(c) Three slope regions delimited by the rising edge of ascending slopes and the
falling edge of descending slopes.

Fig. 7. Routine to compute the array of ascending slopes.

represented by high values in the projection profile, it prevents

against incorrectly dividing a vehicle in two regions.
In steps 3 and 4 of F IND -H ILLS, we call the P HASES routine,
outlined as Fig. 7. This function scans the R and F arrays, with
the order depending on whether we are looking for ascending or
descending regions. In Fig. 8(b), we show the values for these
functions over a sample projection profile.
By pairing the ascending and descending boundaries pro-
duced by F IND -H ILLS in arrays A and D, we can determine
the left and right boundaries of each region of interest, as
exemplified in Fig. 8(c). We assume each region of interest
corresponds to a vehicle in the image. Upper and lower bound-
aries for each region are obtained directly from the binary
segmentation mask M. To guarantee a vehicle’s license plate is
inside its corresponding region of interest, we discard regions
with lower boundaries close to the image bottom—these cases
may correspond to a vehicle that is entering the frame, so its
license plate is not visible yet. Fig. 9. Hypothesis generation for license plate detection. (a) Input image.
(b) Vertical edges. (c) Filtered and dilated edges. (d) Candidate regions.

IV. L ICENSE P LATE D ETECTION

Our detector follows the hypothesis generation and valida-
The motion detector produces one region of interest for tion paradigm [27]. Namely, in the hypothesis generation phase
each moving vehicle present in the scene at a given time. (outlined in Fig. 9) we use edge extraction, edge filtering, and
The license plate detector finds, for each region of interest, region grouping modules to provide coarse candidate regions
an axis-aligned rectangle, which is an approximate bounding based on the edge attribute that makes up the license plate. At
box of the vehicle’s license plate region. This procedure is this phase, we aim to isolate the license plate region and prevent
performed for each region of interest only until a license plate is false negatives, even at the cost of several false positives. In
detected—afterwards, features extracted from the license plate the hypothesis validation phase, we use a region classification
region are tracked across frames, as explained in Section V. module to refine the candidates. For this classification we use
LUVIZON et al.: VIDEO-BASED SYSTEM FOR VEHICLE SPEED MEASUREMENT IN URBAN ROADWAYS 1397

the Text HOG (T-HOG) descriptor [28], which is based on the

observation that the license plate textual information can often
be characterized by the distribution of the directions of the
image gradients.
For edge extraction, we follow the observation from
Zheng et al. [21] that background areas around the license
plate region often have large horizontal edges or small random
noise. Thus, we extract only the vertical image gradients Gx
by convolving the input image with a 3 × 3 horizontal Sobel
operator. The gradients are then stored in a binary edge image
E by comparing each gradient magnitude with the average
magnitude μ of Gx multiplied by some threshold τ (τ = 2 in
our tests), that is

1 if |Gx (x, y)| > μτ
E(x, y) = (3)
0 otherwise
Fig. 10. M AKE -S ET, F IND -S ET, and U NION routines, adapted from
An example of edge extraction is exemplified in Fig. 9(b). As Cormen et al. [32].
the sign of the gradients is not taken into account, this scheme
is invariant to distinctive license plate color schemes.
Edge filtering is performed to remove from E edges that
are too small (below 4 pixels) or too large (above 120 pixels).
This is done using a standard connected component labeling
algorithm and direct comparisons with given minimum and
maximum values for each component’s width and height.
Neighboring vertical edges that remain after filtering are
merged by performing a morphological dilation using a
centered 1 × 7 structuring element. Fig. 9(c) shows an image
example containing filtered and dilated edges.
In order to avoid super-segmenting the license plate, we
group candidate regions according to the geometric criteria
defined by Retornaz and Marcotegui [31]. These criteria take
into account the heights h1 , h2 and widths w1 , w2 of two
bounding boxes b1 and b2 , as well as the coordinates (x1 , y1 )
and (x2 , y2 ) of their centers. Specifically, let h = min(h1 , h2 ),
dx = |x1 − x2 | − (w1 + w2 )/2, and dy = |y1 − y2 |. Then b1
and b2 are said to be compatible — that is, assumed to belong Fig. 11. Region grouping. (a) Region bounding boxes of a sample image.
to the same object — if and only if (b) M AKE -S ET routine applied to all regions, the arrows indicate the node
parent. (c) Result of U NION (w, 7). Adapted from Cormen et al. [32].

|h1 − h2 | < t1 h
dx < t2 h The last stage of our license plate detector is a region
classification step, which discards regions that do not seem to
dy < t3 h (4) contain any textual information. We use for this task the T-HOG
text descriptor [28] which is a texture classifier specialized for
where t1 , t2 and t3 are fixed parameters (respectively, 0.7, 1.1, capturing the gradient distribution characteristic of character
and 0.4 in our tests). strokes in occidental-like scripts. We first estimate a center line
The above criteria are applied to each isolated region by for each candidate image region, by taking, at each column, the
using the union-find data structure, which was adapted from center point between the uppermost and the bottommost pixels
Cormen [32] as shown in Fig. 10. Specifically, at the beginning from the filtered and dilated edge image, as shown in Fig. 12(a)
each region b is a disjoint set created by the M AKE -S ET and (b). Fixed-size windows are then centered at regularly
algorithm, as shown in Fig. 11(a) and (b). The U NION routine spaced points along the center line, as shown in Fig. 12(c). This
then tries to group two compatible candidate regions b1, b2, sub-sampling is done to reduce the computational effort, as well
as shown in Fig. 11(c). These regions are then filtered using as to avoid very similar classifications in neighboring positions.
simple geometric tests that remove regions with dimensions not For each window, we compute the T-HOG descriptor and use
compatible with license plates. In our tests we filtered regions it as an input for a SVM classifier. For the SVM classifier we
with dimensions below 32 × 10 pixels. Fig. 9(d) shows the used a Gaussian χ2 kernel, whose standard deviation parameter
grouped and filtered regions for a sample image. was optimized by cross-validation on a training set of text and
1398 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 18, NO. 6, JUNE 2017

Fig. 12. T-HOG/SVM classification. (a) Candidate region. (b) Region center
line in white color. (c) Sampled points to guide the classification in white
color (step of 4 pixels). (d) and (e) Regions classified as text. (f) Text region
bounding box.

non-text samples as described by Minetto et al. [28].2 The

classifier output is thresholded to give a binary text/non-text
region classification. Fig. 12(d) shows the windows classified
as text for a sample image.
The final license plate detection is performed by taking
regions containing a large number of windows classified as
text [Fig. 12(e)]. Note that vehicles displaying textual labels
and advertisements may have multiple regions satisfying this
Fig. 13. Overview of the proposed feature tracking method.
condition. In those cases, we select the region closest to the
image bottom, which corresponds to the license plate in most
cases. The license plate is approximated by a single axis- be the 2 × 2 gradient matrix in a given window Ω. The region
aligned rectangle that encloses all the text windows from the covered by the window is selected if both eigenvalues of Z are
selected region, as shown in Fig. 12(f). Note that this rectangle above a given threshold. In our system, we used a threshold of 1,
does not have to be accurately adjusted to the license plate—our which leads to a large number of features, but track only the
system will work as expected as long as the feature tracking 10 features with the highest eigenvalues.
module, explained in Section V, can detect enough distinctive The selected features are tracked with subpixel accuracy us-
features inside the rectangle. ing the pyramidal Kanade–Lucas–Tomasi (KLT) [4] algorithm.
A final observation about our detector is that, in our data set, Let I and J be two video frames such that J appears after
the license plates from motorcycles have different dimensions I in the video sequence. The KLT algorithm takes a small
than other vehicles. Moreover, they contain two text lines in- window Ω around each feature extracted from I, and looks
stead of one, and smaller letters and digits, which are more fre- for its corresponding window in J. For a feature centered at
quently merged to each other. Our detector still works for those position u = (xu , yu ), the corresponding window is the one
cases, requiring only adjustments to some thresholds, as well that minimizes the sum of the squared errors, that is
windows spread over the entire region for the T-HOG/SVM  2
classification, instead of following the center line. E= 
I(u) − J(u + d) (6)
Ω

V. F EATURE S ELECTION AND T RACKING where d is the displacement vector that describes the motion
of the feature between frames I and J. To obtain d,  the KLT
Once a license plate region is detected, our system selects
a set of distinctive features and tracks it across multiple video algorithm takes a current estimate e and iteratively solves for
 namely
increments Δd,
frames. Our aim is producing a list containing the trajectories
of the tracked features. The structure of our tracking scheme is   2
E= 
I(u) − J u + (e + Δd) (7)
outlined in Fig. 13.
Feature selection is performed only once for each vehicle, Ω

immediately after its license plate is detected. Following the updating e at each iteration until it converges. As we are
approach from Shi and Tomasi [3], a “good feature” is a region tracking features, the initial estimate for each frame may be the
with high intensity variation in more than one direction, such as displacement d obtained for the previous frame.
textured regions or corners. Let [Ix Iy ] be the image derivatives The traditional Lucas-Kanade algorithm only works for
in the x and y directions of image I, and let small displacements (in the order of one pixel). To overcome
  Ix 2 Ix Iy
 this limitation, we consider the pyramidal version of KLT,
Z= (5) described by Bouguet [33]. The algorithm builds, for each
Ix Iy I2y
Ω frame, a multi-scale image pyramid by using the original image
at the pyramid base, and putting at each subsequent level a
2 We used the source code and training dataset available at www.dainf.ct. version of the image in the previous level with width and height
utfpr.edu.br/%7erminetto/projects/thog.html. reduced by half. The pyramidal KLT algorithm starts by finding
LUVIZON et al.: VIDEO-BASED SYSTEM FOR VEHICLE SPEED MEASUREMENT IN URBAN ROADWAYS 1399

the displacement vector d at the last pyramid level, using the

result as the initial estimate for d in the next level, repeating
the process until the pyramid base (i.e. the original image)
is reached.
The number of levels  in the pyramid determines the
maximum allowed displacement for a feature in pixels, given
by 2+1 −1 (e.g. for  = 3, the maximum displacement is of
15 pixels). Larger values of  may allow for larger
displacements, but they may also produce very small images
Fig. 14. Vehicle speed measurement scheme.
at the upper pyramid levels, which may lead to confusion
and large errors in the initial estimates. Another limitation
of the algorithm is that for the first frame in a sequence (i.e. speed measurement module is converting these measurements
the moment a license plate is detected), there is no known to kilometers per hour (km/h) in the real world.
motion. In these cases, it is common to start with e = (0, 0) in An important assumption of our system is that each street
Equation (7). If there are large feature displacements, e.g. from lane lies on a plane. That assumption makes it possible for us
a vehicle moving at a high speed, this initial estimate will be to map the motion vectors di , which are given in pixels in the
too far from the correct displacement, preventing the feature image plane, to displacements vi , given in meters in the ground
from being properly tracked. plane (see Fig. 14).
To overcome the limitations described above, an initial value Assuming the pinhole camera model, this mapping can be
for d is estimated by using a different method for matching made based on a single view of the scene, through a homog-
features extracted from I and J. We have used the SIFT (Scale- raphy [35] — a plane-to-plane projective transformation—in
Invariant Feature Transform) features proposed by Lowe [5]. a process that is sometimes referred to as inverse perspective
SIFT is a popular method for detecting and describing image mapping [36].
keypoints, being robust to illumination variations and a num- Given a 3 × 3 homography matrix H, an image point pi =
ber of geometric distortions. Using the standard parameters (xi , yi ) can be mapped to the point p̂w = (xw , yw ) in the world
proposed by Lowe [5], SIFT features are extracted from an plane by
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
expanded window around the license plate region from image I, xw zxw xi
and matched to other features extracted from image J, using ⎣ yw ⎦ = ⎣ zyw ⎦ = H ⎣ yi ⎦ . (8)
the nearest neighbor distance ratio matching strategy de- 1 z 1
scribed by Mikolajczyk [34]. The obtained matches can be
used to compute the displacement of each SIFT feature between The homography matrix H may be obtained by associating
frames I and J. The average feature displacement is then four points in the image to known coordinates in the world
used as the initial value for d in Equation (7). Note that this plane. For our tests, we have used as references the markings
process occurs only once for each detected license plate—after left on the asphalt by the inductive loop detectors, which form
the motion is roughly predicted, the system relies on the a rectangle with 4.8 m × 2.0 m (but note that any large enough
KLT algorithm, which allows for faster and more accurate planar object could be used to this purpose). We assume that the
estimates. top-left corner of the rectangle is the origin of the world coor-
The feature selection and tracking module produces, for each dinate system, and use a traditional technique to obtain H [37].
frame, a list of displacement vectors. Since we suppose all Different homography matrices were obtained for each road
the features belong to the same license plate, we expect all lane, i.e. instead of a single ground plane, we assume each lane
the vectors computed for a frame will be similar. Outliers are lies on a different plane.
rejected using an iterative method: for each set of displacement The output of the feature selection and tracking method, for
vectors, we compute the mean and standard deviation, and each vehicle and each pair of frames, is a set of motion vectors
remove from the set those vectors outside the three-sigma di = ui (t) − ui (t − Δt), where ui (t) is the feature position
deviation, with the process being repeated until the standard in the current video frame, ui (t − Δt) is the feature position
deviation is smaller than 0.5 pixel. The outlier rejection routine in the previous video frame, Δt is the frame interval, and
is used independently for the x and y directions, and applied i = {1, 2, . . . , n} is a sequence of tracked features. We compute
both to the vectors computed by the KLT algorithm and the the feature displacements in the real world, denoted by vi , from
initial estimates obtained from matching SIFT features. Equation (8), namely

vi = Hui (t) − Hui (t − Δt). (9)

VI. S PEED M EASUREMENT The displacement in meters between frames t − Δt and t

for a motion vector ui can be obtained by vi , the Euclidean
Our system measures vehicle speeds based on the motion
norm of vi . Each displacement vector can be associated with a
vectors obtained by the feature selection and tracking method.
measurement of the vehicle’s instantaneous speed, given by
Each motion vector di can be associated with a measurement
of a vehicle’s instantaneous speed at a particular time, given vi 
si = (10)
in pixels per frame, in the image plane. The purpose of the Δt
1400 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 18, NO. 6, JUNE 2017

TABLE I
D ATA S ET I NFORMATION : T IME (M INUTES ); N UMBER OF V IDEOS ;
N UMBER OF V EHICLES W ITH P LATES AND S PEED I NFORMATION .
T HE Q UALITY O PTIONS A RE AS F OLLOWS : [H]—H IGH Q UALITY;
[N]—F RAMES A FFECTED BY N ATURAL OR A RTIFICIAL N OISE ;
[L]—F RAMES A FFECTED BY S EVERE L IGHTING C ONDITIONS ;
[B]—M OTION B LUR ; AND [R]—R AIN

Fig. 15. Vehicle speed distribution.

speeds were obtained from a high precision speed meter based

on inductive loop detector, properly calibrated and approved
by the Brazilian national metrology agency (Inmetro). Note
that the videos contain some vehicles with no visible license
where Δt is the time, in seconds, between two frames. This plate, and that the ground truth speed meter sometimes fails to
time is supposed to be constant, and is the inverse of the frame properly assign a speed to a vehicle. The “No. valid” column
rate—e.g. for a frame rate of 30 frames per second, Δt = 1/30. in Table I indicates the number of vehicles which have both a
The instantaneous vehicle speed s is estimated by averaging the visible license plate and an assigned speed.
values of si for a set of tracked features. Fig. 15 shows how the vehicle speeds are distributed (the
The assumption that all the motion vectors for a vehicle speed limit in this particular roadway is 60 km/h).
lie on the same plane is a simplification, used so that the In our manual ground truth annotation, we also include a flag
actual 3D position of the tracked features does not have to be for motorcycles and non-motorcycles (ordinary vehicles). We
discovered. As the actual license plates are always above the noted that motorcycles pose a challenge for the ground truth
road level, the computed speeds will be higher than the actual speed meter, which was able to measure the speed in only
speeds. To mitigate the effects of these erroneous measurements 43% of the cases (compared to 92% for ordinary vehicles). For
we multiply each vehicle’s measured speed by a constant that reason, motorcycles represent 4.5% of the total number of
factor S, which we set to 0.9 in our experiments. As shown vehicles, but only 2.1% of the “No. valid” vehicles.
in Section VII, the use of the S factor is simple but effective, The whole dataset used in our experiments will be made
as long as the tracked features are at approximately the same available for research purposes, and can be itself considered one
distance from the ground. of the major contributions of our work.
The final speed for a vehicle is obtained by averaging the
instantaneous speed across multiple frames, while the vehicle
is located at a certain image region. In our experiments, we B. Motion Detection Evaluation
considered a speed measurement region close to the ground To evaluate the results from the motion detector, we compare
truth loop detectors, to allow a direct comparison between the the obtained regions of interest (ROIs) with the license plates
measured speeds and the ground truth speeds. in the ground truth. Ideally, all the detected ROIs will contain
a license plate, and all license plates will be fully contained
VII. E XPERIMENTS within a detected ROI. Objectively, we compute precision and
recall metrics as described by Wolf et al. [38], with the precision
A proof-of-concept system was built for evaluating the pro- p being given by the proportion of ROIs which contain at least
posed approach. Besides the cameras and physical infrastruc- one license plate, and the recall r being given by the proportion
ture, we used a 2.2 GHz Intel Core i7 machine with 12 GB of of license plates that were inside a ROI. Namely, we compute
RAM running Linux, with the algorithms implemented in C++.
|D|
 |G|

In the next sections we describe our dataset, and evaluate our
m(di , G) m(gi , D)
system’s performance regarding motion detection, license plate i=1 i=1
detection, and speed measurement. p= r= (11)
|D| |G|

where G = {g1 , g2 , . . . , g|G| } is the set of ground truth license

A. Dataset plate regions, and D = {d1 , d2 , . . . , d|D| } is the set of detected
Our dataset, summarized in Table I, contains 20 videos cap- ROIs. Function m is defined by
tured by a low-cost 5-megapixel CMOS image sensor, with
frame resolution of 1920 × 1080 pixels, at 30.15 frames per m(a, S) = max m (a, si ) (12)
i={0,...,|S|}
second. The videos are divided in 5 sets according to weather
and recording conditions. Each video has an associated ground where m is a function that compares two rectangular regions a
truth file, in a simple XML format, containing bounding boxes and b — more specifically, a ROI with a license plate region:
for the first license plate occurrence of each vehicle, as well as  area(a∩b)
each vehicle’s actual speed. The ground truth for the license  1 if min(area(a),area(b)) >λ
m (a, b) = (13)
plates was obtained by human inspection. The ground truth 0 otherwise
LUVIZON et al.: VIDEO-BASED SYSTEM FOR VEHICLE SPEED MEASUREMENT IN URBAN ROADWAYS 1401

TABLE II
M OTION D ETECTION P ERFORMANCE : THE P RECISION p, R ECALL r, AND
AVERAGE T IME ( IN M ILLISECONDS , FOR E ACH F RAME ) FOR
F IVE S UBSAMPLING C ONFIGURATIONS AND
T WO OVERLAPPING T HRESHOLDS

TABLE III
L ICENSE P LATE D ETECTION P ERFORMANCE E VALUATION , BASED ON
P RECISION (p), R ECALL (r), AND THE F -M EASURE . T HE VALUES IN
B OLDFACE A RE THE M AXIMA O BTAINED FOR E ACH C ASE

with the λ threshold indicating how much of the license plate Fig. 16. Examples of license plates detected by our system, for representative
region must be contained within the ROI. We performed tests samples of each set.
using two different values for λ: 1.0 (i.e. the entire license plate
is contained within the ROI) and 0.5.
Table II shows the precision and recall, as well as the average
processing time, obtained by our motion detector. The different
columns show the results obtained with different amounts of
subsampling — more sparse grids will reduce the processing
time, but can also lead to incorrect results. Fig. 17. Examples of license plates not detected by our system.

C. License Plate Detection Evaluation results obtained for the entire data set are shown in Table III,
To evaluate the performance of the license plate detector, we divided in 5 subsets according to weather and recording
compare the detected license plates with those in the ground conditions. Our detector significantly outperformed the other
truth. The comparison is based on the same precision and approaches in these tests. The average time to process each
recall metrics used for evaluating the motion detector (see region of interest was 58 ms for SnooperText; 918 ms for
Section VII-B), with some differences. First, set D refers to the Zheng et al.; 402 ms for SWT; and 195 ms for our detector.
set of detected license plates. Second, function m is defined as Examples of license plates detected by the proposed method
in the PASCAL Visual Object Detection Challenge [39]: are shown in Fig. 16. Our detector worked as expected even
 in some situations with severe image noise or motion blur.
1 if area(a∩b)
area(a∪b) > 0.5 Detection errors occurred mainly in the hypothesis generation
m (a, b) = (14)
0 otherwise. phase, with true license plate regions being eliminated by
some filtering criteria when they became connected with a
For ranking purposes, we also consider the F -measure, background region. Samples of license plates not detected by
which is the harmonic mean of precision and recall: F = 2 · our system are shown in Fig. 17.
p · r/(p + r).
We compared our license plate detector with three text
D. Vehicle Speed Measurement Evaluation
and license plate detectors described in the literature (see
Section II-B): SnooperText [27], the Zheng et al. [21] algo- Speed measurement performance was evaluated by compar-
rithm, and the Stroke Width Transform (SWT) [22]. The param- ing the speeds measured by our system with the ground truth
eters for these detectors were obtained by running the system speeds obtained by the inductive loop detectors. According to
on 25% of the videos from the dataset, and selecting the param- the standards adopted in the USA, an acceptable measurement
eter combinations that produced the highest F-measures. The must be within the [−3 km/h, +2 km/h] error interval.
1402 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 18, NO. 6, JUNE 2017

TABLE IV allowed error interval decreased significantly. That happens

S PEED M EASUREMENT R ESULTS O BTAINED BY O UR S YSTEM AND
O THER A PPROACHES : “L OWER ,” “I DEAL ,” AND “H IGHER ” R EPRESENT because in this case the features have a larger variance in their
S PEED E RRORS B ELOW, A BOVE , AND W ITHIN THE A CCEPTABLE L IMITS , heights from the ground—thus, motion vectors computed from
R ESPECTIVELY, C ONSIDERING THE U.S. S TANDARD [−3/+2 KM / H ] the same vehicle have very different lengths. Moreover, these
features are less distinctive, leading to tracking errors.
We also tested our system with an “ideal license plate de-
tector,” taking as references the manually annotated license
plates from the ground truth, instead of the detected license
plates. That includes some license plate regions that are hard
to identify even by a human observer. As shown in the third
row of Table IV, the performance in this case was not much
different from the performance obtained by our complete sys-
tem, indicating that the tracking can be done even with poor
license plate regions, hard to be identified even by an human
observer.
We also compared our system with a blob-based tracker. In
this experiment, we used a particle filter algorithm [40] to track
the regions of interest found by our motion detection module.
We tested several parameter combinations for this approach,
but we were unable to find a suitable configuration for our
application. We believe that the main reason for this is that our
camera is installed very close to the vehicles, in such a way that
a probabilistic search cannot precisely define the position of the
vehicle in all frames. Table IV shows the best results we could
obtain using the blob tracking approach.

VIII. C ONCLUSION
Fig. 18. Speed measurement error distribution. This paper addressed the problem of measuring vehicle
speeds based on videos captured in an urban setting. We pro-
posed a system based on the selection and tracking of distinctive
The first row in Table IV shows the results obtained features located within each vehicle’s license plate region. The
by our system. Percentages are given regarding the valid system was tested on almost five hours of videos with full-
vehicles—those with both a license plate and an associated HD quality, with more than 8,000 vehicles in three different
speed in the ground truth—and are divided in 3 classes, depend- road lanes, with associated ground truth speeds obtained by
ing on whether the measured speed was below, inside, or above a high precision system based on inductive loop detectors, as
the acceptable error interval. Fig. 18 shows the distribution well as manually labeled ground truth license plate regions. Our
of the measurement errors, with 96% of the measurements system uses a novel license plate detection method, based on a
being inside the acceptable limits. The maximum nominal texture classifier specialized to capture the gradient distribution
error values for the whole dataset were −4.68 km/h and characteristics of character strokes that make the license plate
+6.00 km/h, with an average of −0.5 km/h a standard deviation letters. This module achieved a precision of 0.93 and a recall
of 1.36 km/h. We observed that the assumption that all the of 0.87, outperforming other well-known approaches. We have
license plates have nearly the same distance from the ground also shown that extracting distinctive features from the license
is the main cause of speed measurement errors: when the plate region led to better results than taking features spread over
license plates are very high above the ground (e.g. in buses the whole vehicle, as well as an approach which uses a particle
or trucks) the measured speed can be higher than the actual filter for blob tracking. In our experiments, the measured speeds
speed, with the opposite occurring when the license plates had an average error of −0.5 km/h, staying in over 96.0% of the
are unusually low. A total of 99.2% of the vehicles were cases inside the +2/−3 km/h error interval determined by the
successfully tracked until they reached the speed measurement regulatory authorities in several countries.
region. On average, our tracking module spent 49.8 millisec- As future work, we intend to verify if estimating the distance
onds per frame. Examples of measured speeds are shown of the license plates from the ground can improve the results.
in Fig. 19. We also aim to apply an OCR on the detected license plates
In order to verify if distinctive features from a license plate in order to create a traffic speed control system with integrated
region are a good choice for measuring a vehicle’s speed, we surveillance tools, e.g. to compute the traffic flow, to identify
performed tests using a version of our system which takes fea- stolen vehicles, etc. Another topic for future work is the imple-
tures from the whole vehicle region (“free feature selection”). mentation on a compact platform that allows local processing,
The results are shown in Table IV. It can be seen that the including optimizations such as parallel processing on GPUs,
percentage of vehicles whose measured speed is inside the thus reducing communication bandwidth requirements.
LUVIZON et al.: VIDEO-BASED SYSTEM FOR VEHICLE SPEED MEASUREMENT IN URBAN ROADWAYS 1403

Fig. 19. Examples of vehicle speeds measured by our system and by a high-precision meter based on inductive loops.

R EFERENCES [10] C. H. Xiao and N. H. C. Yung, “A novel algorithm for estimating vehicle
speed from two consecutive images,” in Proc. IEEE WACV, 2007, pp. 1–6.
[1] T. V. Mathew, “Intrusive and non-intrusive technologies,” Indian Inst. [11] H. Zhiwei, L. Yuanyuan, and Y. Xueyi, “Models of vehicle speeds
Technol. Bombay, Mumbai, India, Tech. Rep., 2014. measurement with a single camera,” in Proc. Int. Conf. Comput. Intell.
[2] N. Buch, S. Velastin, and J. Orwell, “A review of computer vision tech- Security Workshops, 2007, pp. 283–286.
niques for the analysis of urban traffic,” IEEE Trans. Intell. Transp. Syst., [12] C. Maduro, K. Batista, P. Peixoto, and J. Batista, “Estimation of vehicle
vol. 12, no. 3, pp. 920–939, Sep. 2011. velocity and traffic intensity using rectified images,” in Proc. IEEE ICIP,
[3] J. Shi and C. Tomasi, “Good features to track,” in Proc. IEEE Int. Conf. 2008, pp. 777–780.
CVPR, 1994, pp. 593–600. [13] H. Palaio, C. Maduro, K. Batista, and J. Batista, “Ground plane
[4] B. D. Lucas and T. Kanade, “An iterative image registration technique velocity estimation embedding rectification on a particle filter multi-target
with an application to stereo vision,” in Proc. Joint Conf. Artif. Intell., tracking,” in Proc. IEEE ICRA, 2009, pp. 825–830.
1981, pp. 674–679. [14] L. Grammatikopoulos, G. Karras, and E. Petsa, “Automatic estimation
[5] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” of vehicle speed from uncalibrated video sequences,” in Proc. Mod.
Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, 2004. Technol., Educ. Prof. Pract. Geodesy Related Fields, 2005, pp. 332–338.
[6] D. Luvizon, B. Nassu, and R. Minetto, “Vehicle speed estimation by [15] T. Schoepflin and D. Dailey, “Dynamic camera calibration of roadside
license plate detection and tracking,” in Proc. IEEE ICASSP, 2014, traffic management cameras for vehicle speed estimation,” IEEE Trans.
pp. 6563–6567. Intell. Transp. Syst., vol. 4, no. 2, pp. 90–98, Jun. 2003.
[7] D. Dailey, F. Cathey, and S. Pumrin, “An algorithm to estimate mean [16] G. Garibotto, P. Castello, E. Del Ninno, P. Pedrazzi, and G. Zan, “Speed-
traffic speed using uncalibrated cameras,” IEEE Trans. Intell. Transp. vision: Speed measurement by license plate reading and tracking,” in
Syst., vol. 1, no. 2, pp. 98–107, Feb. 2000. Proc. IEEE Trans. Intell. Transp. Syst., 2001, pp. 585–590.
[8] V. Madasu and M. Hanmandlu, “Estimation of vehicle speed by motion [17] W. Czajewski and M. Iwanowski, “Vision-based vehicle speed mea-
tracking on image sequences,” in Proc. IEEE Intell. Veh. Symp., 2010, surement method,” in Proc. Int. Conf. Comput. Vis. Graphics, 2010,
pp. 185–190. pp. 308–315.
[9] S. Dogan, M. S. Temiz, and S. Kulur, “Real time speed estimation of mov- [18] M. Garg and S. Goel, “Real-time license plate recognition and speed
ing vehicles from side view images from an uncalibrated video camera,” estimation from video sequences,” ITSI Trans. Electr. Electron. Eng.,
Sensors, vol. 10, no. 5, pp. 4805–4824, 2010. vol. 1, no. 5, pp. 1–4, 2013.
1404 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 18, NO. 6, JUNE 2017

[19] C.-N. E. Anagnostopoulos, I. E. Anagnostopoulos, I. D. Psoroulas, [37] D. G. R. Bradski and A. Kaehler, Learning OpenCV, 1st ed. Sebastopol,
V. Loumos, and E. Kayafas, “License plate recognition from still images CA, USA: O’Reilly Media, 2008.
and video sequences: A survey,” IEEE Trans. Intell. Transp. Syst., vol. 9, [38] C. Wolf and J.-M. Jolion, “Object count/area graphs for the evaluation
no. 3, pp. 377–391, Mar. 2008. of object detection and segmentation algorithms,” Int. J. Doc. Anal.
[20] S. Du, M. Ibrahim, M. Shehata, and W. Badawy, “Automatic License Plate Recognit., vol. 8, no. 4, pp. 280–296, 2006.
Recognition (ALPR): A state-of-the-art review,” IEEE Trans. Circuits [39] M. Everingham, L. V. Gool, C. K. I. Williams, J. Winn, and A. Zisserman,
Syst. Video Technol., vol. 23, no. 2, pp. 311–325, Feb. 2013. “The PASCAL Visual Object Classes (VOC) challenge,” Int. J. Comput.
[21] D. Zheng, Y. Zhao, and J. Wang, “An efficient method of license plate Vis., vol. 88, pp. 303–338, 2009.
location,” Pattern Recognit. Lett., vol. 26, no. 15, pp. 2431–2438, 2005. [40] P. Perez, C. Hue, J. Vermaak, and M. Gangnet, “Color-based probabilistic
[22] B. Epshtein, E. Ofek, and Y. Wexler, “Detecting text in natural scenes tracking,” in Proc. ECCV, 2002, pp. 661–675.
with stroke width transform,” in Proc. IEEE Int. Conf CVPR, 2010,
pp. 886–893.
[23] B. Li, B. Tian, Y. Li, and D. Wen, “Component-based license plate detec-
tion using conditional random field model,” IEEE Trans. Intell. Transp. Diogo Carbonera Luvizon received the M.Sc. de-
Syst., vol. 14, no. 4, pp. 1690–1699, Dec. 2013. gree from the Federal University of Technology of
[24] A. Ashtari, M. Nordin, and M. Fathy, “An Iranian license plate recogni- Paraná (UTFPR), Curitiba, Brazil, in 2015. He is cur-
tion system based on color features,” IEEE Trans. Intell. Transp. Syst., rently working toward the Ph.D. degree at Université
vol. 15, no. 4, pp. 1690–1705, Aug. 2014. de Cergy-Pontoise, Cergy-Pontoise, France.
[25] H. Li, D. Doermann, and O. Kia, “Automatic text detection and tracking From 2010 to 2014, he was a Research Engi-
in digital video,” IEEE Trans. Image Process., vol. 9, no. 1, pp. 147–156, neer in a company of vehicle speed measurement
Jan. 2000. systems. His main research interests include vehicle
[26] W. Zhou, H. Li, Y. Lu, and Q. Tian, “Principal visual word discovery for detection and tracking, speed estimation, and image
automatic license plate detection,” IEEE Trans. Image Process., vol. 21, descriptors.
no. 9, pp. 4269–4279, Sep. 2012.
[27] R. Minetto, N. Thome, M. Cord, N. J. Leite, and J. Stolfi, “SnooperText:
A text detection system for automatic indexing of urban scenes,” Comput.
Vis. Image Understand., vol. 122, pp. 92–104, 2014. Bogdan Tomoyuki Nassu received the Ph.D. degree
[28] R. Minetto, N. Thome, M. Cord, J. Stolfi, and N. J. Leite, “T-HOG: An in advanced interdisciplinary studies from The Uni-
effective gradient-based descriptor for single line text regions,” Pattern versity of Tokyo, Tokyo, Japan, in 2008.
Recognit., vol. 46, no. 3, pp. 1078–1090, 2013. From 2008 to 2011, he was a Researcher with
[29] A. Bobick and J. Davis, “The recognition of human movement using the Railway Technical Research Institute, Tokyo, and
temporal templates,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 23, a Postdoctoral Research Fellow with the Federal
no. 3, pp. 257–267, Mar. 2001. University of Parana, Curitiba, Brazil. Since 2012, he
[30] J. Ha, R. Haralick, and I. Phillips, “Document page decomposition by the has been an Assistant Professor with the Federal Uni-
bounding-box project,” in Proc. ICDAR, 1995, vol. 2, pp. 1119–1122. versity of Technology of Paraná (UTFPR), Curitiba.
[31] T. Retornaz and B. Marcotegui, “Scene text localization based on the His main research interest is applying computer
ultimate opening,” in Proc. ISMM, 2007, vol. 1, pp. 177–188. vision techniques to practical problems.
[32] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction to
Algorithms, 3rd ed. Cambridge, MA, USA: MIT Press, 2009.
[33] J.-Y. Bouguet, “Pyramidal Implementation of the Lucas Kanade Feature
Tracker,” Intel Corp., Microprocessor Res. Lab., Mountain View, CA, Rodrigo Minetto received the Ph.D. degree in
USA, 2000. computer science from Université Pierre et Marie
[34] K. Mikolajczyk and C. Schmid, “A performance evaluation of local Curie, Paris, France, and the University of Campinas,
Campinas, Brazil, in 2012.
descriptors,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 10,
pp. 1615–1630, Oct. 2005. Since 2012, he has been an Assistant Professor
[35] G. Wang, Z. Hu, F. Wu, and H.-T. Tsui, “Single view metrology from with the Federal University of Technology of Paraná
(UTFPR), Curitiba, Brazil. His main research inter-
scene constraints,” Image Vis. Comput., vol. 23, no. 9, pp. 831–840, 2005.
[36] H. Li, M. Feng, and X. Wang, “Inverse perspective mapping based urban ests include text detection, object tracking, image
road markings detection,” in Proc. IEEE Int. Conf. CCIS, Oct. 2012, descriptors, and additive manufacturing.
vol. 3, pp. 1178–1182.