=Paper= {{Paper |id=Vol-2691/paper56 |storemode=property |title=Light Invariant Lane Detection Method Using Advanced Clustering Techniques |pdfUrl=https://ceur-ws.org/Vol-2691/paper56.pdf |volume=Vol-2691 |authors=Aleksandr Karavaev,Rami I. Al-Naim }} ==Light Invariant Lane Detection Method Using Advanced Clustering Techniques== https://ceur-ws.org/Vol-2691/paper56.pdf

Light Invariant Lane Detection Method Using
Advanced Clustering Techniques

Aleksandr Karavaev Rami Al-Naim
ITMO University ITMO University
Saint-Petersburg, Russia Saint-Petersburg, Russia
alexkaravaev@protonmail.com rami.naim2010@yandex.ru

Abstract—In this paper we propose a novel approach to de- Tesla company shows, that only-camera solution is possible2 .
tecting road lanes from video stream in bad-light road scenarios. Main disadvantage of lidar is its price, that is comparable in
The main focus of this article is given to the introduction new some cases to the price of the car itself. And most of the
image binarization method in non-common color space followed solutions require a lot more than one lidar on the car (up to 6
by improved density hierarchical clustering algorithm called small and big lidars, that are mounted on various sides of the
HDBSCAN. These techniques allow to detect lane boundaries
car). Some researches and forecasters insist that the price of
even in low-light scenarios with robust and parameter-free setup.
lidars will eventually fall down3 . However, other ones compare
camera with human visual cortex, which can reliably identify
Keywords—Lane detection, Computer vision, HDBSCAN, DB- and detect distance to various objects4 .
SCAN, Image Thresholding
And this is our motivation why we focus on developing
system, that detects road lane markings with the cameras.
I. I NTRODUCTION We believe that such a solution should meet the following
requirements:
The aim of this paper is to design low-cost in terms of
processing time pipeline for robust road marking detection in • System should be cheap for maintaining and produc-
low-level of light road scenarios. ing for the purpose of high scalability.
Self-driving cars are actively introduced into people lives. • System should be robust to rapidly changing light
Their number and complexity of software of on-board com- environment.
puters are increasing [1]. Cars of only one Waymo company
managed to drive twenty million miles in self-driving mode1 . • System should be fast for not powerful computers and
Nowadays most of companies are relying on solutions with shouldn’t require a lot of computational power.
deep neural networks for perception module of the car [2]. Main proposals of our paper:
Moreover most of them are using radars and lidars, which
allow to perceive the environment even in dim and dark road • Image processing in color space CIE L*a*b, that is
scenarios unlike to usual camera [3]. Despite the fact that decreasing light impact on the scene and image.
these approaches are dominant in the field, they have major
limitations that restrict them from full implementation in the • New formula for image binarization.
industry and interfere with scalability to more users [4]. • Using HDBSCAN — more recent method of density
There is a need for training deep neural networks on clustering instead of DBSCAN.
a powerful machine equipped with a lot of video cards • Using color information in clustering.
for achievement of good precision and recall of an output
detections. Furthermore, even after successful training and Combined together, these methods provide a robust
deploying such big model developers need to install high pipeline with few parameters that need to be configured. Thus,
performance computers on the car because real-time execution the time for configuring is reduced, which seems very useful.
is essential in the case of self-driving car. This solution is By the term pipeline here and after we mean a set of data
more difficult to scale, not to mention the trend to small-size processing elements connected in series, where the output of
components and reducing their cost. Interest in single-board one element is the input of the next one5 .
computers is gradually rising because their best qualities — 2 Tesla official website, autopilot description,
compactness and price [5]. https://www.tesla.com/autopilot
3 A. Davies, ”This Lidar Is So Cheap It Could Make Self-Driving a Real-
A lot of disputes in the community of self-driving cars are
ity”, 7 Nov 2019, https://www.wired.com/story/lidar-cheap-make-self-driving-
ongoing right now. Some researchers consider advantages and reality/
disadvantages of using cameras or lidars [6]. The experience of 4 B. Templeton, ”Elon Musk’s War On LIDAR:
Who Is Right And Why Do They Think That?”,
1 K. Wiggers, ”Waymo’s autonomous cars have driven 20 million miles https://www.forbes.com/sites/bradtempleton/2019/05/06/elon-musks-war-
on public roads”, 6 Jan 2020, https://venturebeat.com/2020/01/06/waymos- on-lidar-who-is-right-and-why-do-they-think-that/#5b6971232a3b
autonomous-cars-have-driven-20-million-miles-on-public-roads/ 5 Pipeline (Computing), https://en.wikipedia.org/wiki/Pipeline (computing)

Copyright© 2020 for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
The paper is organized as follows. Section 2 is giving C. Using Clustering For Denoising
brief review of other papers in the field. Section 3 describes
Some road lane detection methods use density clustering
proposed approach in details. Section 4 compares proposed
for finding lane on the road image. For example, authors in
approach with other ones. Finally, Section 5 gives the conclu-
paper [13] use similar approach for clustering points belonging
sion.
to lane marking. However, they use pretty old and outdated
clustering algorithm DBSCAN. Besides that, authors use dif-
II. R ELATED W ORK ferent threshold operation for the initial steps of processing the
A. Neural Network approach image and Otsu binarization [14], which can fail or give bad
result in some cases of road scenarios.
In recent years neural networks have become a common
tool in image processing tasks. In the field of driving an Aside from this paper, clustering in the lane marking sce-
autonomous car, neural networks are often used to detect road nario is used in the work [15]. In this paper authors use simple
markings, obstacles and road signs. The article [7] describes hierarchical clustering and new method of post-processing
a method for detecting road marking lines based on neural clustering results. For every calculated cluster they calculate
networks. The authors provide experimental data indicating the slope line of it, after that they count the intersections of
a high accuracy of detection of road marking lines and a these slopes with other clusters. The clusters with the most
high image processing speed. However, the experiments were number of intersections are used later. This improvement deals
conducted on equipment, the cost of which is approximately with filtration of clusters that are elongated more on the y axis
equal to $2,000, and such a price may not be acceptable. The instead of x axis and utilized the geometrical feature of almost
article [8] presents experimental data for several algorithms for every road marking, which are located mostly one on the top of
detecting lanes using neural networks. All algorithms presented another on one line. This improvement work well on straight
in this article use either expensive or specialized equipment road scenarios, nonetheless it can fail on turns or curb road
for image processing. This can lead to a significant increase environment.
in cost, or narrow the scope of the possible application of the
system. III. ROAD M ARKING D ETECTION
In this section we examine our algorithm in details. The
B. Binarization and Processing in a Different Color Spaces full block diagram of the approach can be found on Fig. 2
Materials used for carriageway marking lines often have Data: Input video stream
bright colours which differ a lot from the rest the road’s Result: Clustered image points
surface. Consequently, some algorithms for road marking line initialization;
detection utilize this quality and use prior knowledge of while not end of video do
colours for thresholding [9], [10]. For binarization based on Read current color frame;
beforehand estimated colours’ thresholds the HSV color space Convert image into CIE Lab color space;
is widely used. It shows a good results, however, with different Normalize L-channel using MIN-MAX method;
illumination of a scene the colors detected by the camera Calculate mean and standard deviation of the
are different [11]. Thus, with changes in the illumination the L-image;
chromatic values of the pixels on the image may also differ, Calculate threshold value;
which leads to a significant amount of noise on the binarized Binarize L-channel;
image (Fig. 1). As a result, the further applied algorithms for Image < − Select pixels from initial color frames,
lane detection, such as Hough transformation [12], may not where thresholded image != 0;
give an expected result and fail in finding the lanes. Such Initalize HDBSCAN with selected parametrs;
behavior can be handled by setting the algorithm’s parameters, Downscale image;
but this often decreases robustness and reduces number of Cluster Labels, Labels Probabilites < − Clusterize
scenarios in which the algorithm can be used. image with HDBSCAN algorithm;
Discard labels with low probability(< 0.75);
Mask labels to image;
Resize image with labels to initial size;
end
Algorithm 1: Full proposed algorithm

A. Binarization in CIELab Color Space
It is necessary to define two assumptions in order to choose
a threshold value of binary mask with road marking from an
image.
Assumption 1: Road marking lines takes about 5% of an
image.
Fig. 1. Example of a poor binarized mask: color of road lines in HSV color
space changed due to changes in illumination
Assumption 2: Road marking lines brighter than majority
of objects on a road image (road surface, roadside, cars, etc.).
Pixels of road marking lines on the image are located at
Input video the rightmost side of the normalized histogram (5% of the
stream whole distribution). In order to choose appropriate threshold
value the property of a Gaussian normal distribution and 3σ
rule is used [19]. We calculate the mean µ and standard
deviation σ for the normalized histogram of the image and
use Gaussian distribution to estimate the threshold value. This
Converting to CIE lab value bounds 5% of the distribution of the brightest pixels
colorspace which represent road marking lines. Example of approximation
of a normalized histogram can be seen at Fig. 3, where red
line represents normal distribution. The following equation is
Color image used for threshold value estimation:
Adaptive σ
thresholding t = µ + σ(k +
), (1)
2σu
where t — threshold value; µ — mean value from normalized
histogram; σ — standard deviation from normalized histogram;
σu — standard deviation of an uniform distribution; k —
scaling coefficient.
Combining thresholded
binary mask and color mask Histogram Normalized Histogram

0.04 0.04

Probability

Probability
0.02 0.02
HDBSCAN hierarchical
clustering 0.00 0.00
0 50 100 150 200 250 0 50 100 150 200 250
Illumination level Illumination level
Image

Result
Fig. 3. Example of histogram normalization and its approximation with
Fig. 2. Total diagram of our approach Gaussian distribution

From (1) it is clear that the threshold value lies in the
The image with road segment is captured by camera in half-interval (2σ; 3σ] and the exact value depends on standard
RGB color space. In order to reduce noise on the image deviation. The standard deviation of the uniform distribution
we preprocess the image by applying Gaussian blur [16]. In is used to normalize the standard deviation of the histogram
experiments the kernel with size 15 was chosen to filter high- and to accurately determine where the threshold value lies
frequency noises on the image. between the mentioned interval. The scaling parameter k is
chosen depending on how much area of an image is covered
The next step is converting the image from RGB to by road marking lines. For real road application we propose
desired color space. We propose to use one of the perceptually to use k = 2. To illustrate the interval in which the threshold
uniform color space — CIELab [17]. In this color space each value lies the image from Duckietown Project is used (Fig. 4)
pixel is encoded with three values: L, a and b. L describes [20]. Since road marking lines on this image cover a larger
the brightness of a pixel, in other words, characteristic of area compared to real road images, scaling coefficient is set to
luminance. a and b values describe chromatic characteristics, 1. Group of pixels on the right side of normalized histogram
from green to red and from blue to yellow, respectively. represent road marking lines. They are inside of the threshold
marked by a rectangle on the plot.
Further, we propose to calculate histogram of the L channel
of the image in CIELab color space and normalize it. Con- If a scene of image is well-illuminated, its contrast is high.
sidering aforementioned assumptions, pixels of road marking Thus, values on the histogram that correspond to the pixels of
lines are located in the upper right part of the distribution. For the road will be located near each other at the left side of the
the histogram equalization the max-RGB like method is used distribution, while values with intensities of a road marking
[18]. This approach utilize the fact that humans perceive color lines will be located at the most right part of the histogram.
relative to the contrast of the full image, i.e. difference between In that case standard distribution will be rather small and the
brightest and darkest point of the image. As a result, the ratio of σ/σu consequently also will be small. Because of this
histogram is normalized in such a way that its low boundary threshold value will be close to 2σ, which guarantees that it
is a minimum value of a non-zero value of a brightness of the will be less than values of the pixels of bright road marking
image in L channel, and upper boundary is maximum value lines. Contrariwise, when the scene of image is dark, histogram
of a brightness of the image in L channel. might not have distinct groups of pixels of the road or road
marking lines. The standard deviation of such distribution will In this paper we selected HDBSCAN [21] as the main
be greater than in the case described before, so the ratio of clustering algorithm which is declared as an improved hierar-
σu will be relatively large. As a result, calculated threshold chical version of more older algorithm DBSCAN [22]. Main
σ

value will be close to 3σ, so it might be grater than values of improvements over DBSCAN are as follows:
the road marking lines, but it reduces amount of fake pixels
marked as road marking lines (pixels of background or road • The algorithm has much fewer parameters that need to
itself). be configured, because during data processing HDB-
SCAN selects the best parameters according to its own
The results of image binarization using described algorithm indicators, for example, the epsilon parameter.
can be seen in Fig. 5. As a method for comparison Otsu
binarization was chosen because it is one of the most pop- • The algorithm can find clusters with densities varying
ular method of threshold value calculation. Our approach for within the cluster area.
threshold value calculation shows better results: binary mask • The clustering procedure not only assigns a cluster
contains less noise and registers even yellow lines of road number to each input point, but also calculates vector
marking. This is possible because of those two assumptions of probabilities where each probability reflects how
at the beginning of this subsection. probable a point belongs to each cluster or noise.
Now it is worth to emphasize why the main advantages of
B. Hierarchical Clustering
the clustering algorithm are critically important in lane finding
After the initial processing stage (binarization) it is nec- scenario. Even on one route a car might encounter different
essary to obtain information from the black-and-white binary light conditions, from completely dim (e.g tunnel) to sunny and
image about which pixels belong to road marking and which bright, therefore, there is a need to develop a robust pipeline
are just road or other miscellaneous noise or even error from with several hyper-parameters to configure.
binarization algorithm used before. The proposed algorithm is The quality of road markings in reality might be poor
as follows. and the markings might be partially erased, therefore, it is
Given that we have the black and white binary mask and necessary to be able to detect road markings with holes inside
the original color image, we combine them into one resulting and at the same time to detect them as single cluster.
image. Pixels that were previously white receive color from In the task of autonomous driving the safety comes first, so
the original image, and pixels that were plain black remain we select only the points with the lowest probability of noise.
black. The result of this step of the pipeline is showed at Fig. 6.
This procedure is aimed to improve the next clustering
part. This increases the useful information for the algorithm, IV. R ESULTS
because we will have not only information about the intensity
of the pixels, but also the color, and therefore, it is easier to A. Comparison with Other Approaches
divide the points into meaningful clusters. For example, we can The Fig. 7 represents visual part of comparison of the
divide into separate groups the road lane and asphalt, which clustering algorithms. The parameters for clustering algorithms
surrounds lane on all sides but has a completely different RGB were chosen as follows:
color.
• For K-Means:
Histogram Normalized Histogram ◦ n clusters = 6,
0.04 0.04 ◦ random state = 0,

0.03 0.03
• For DBSCAN:
Probability

Probability

◦ eps = 10,
0.02 0.02 ◦ min samples = 200,
0.01 0.01 • For HDBSCAN:
◦ min cluster size = 500,
0.00 0.00
0 50 100 150 200 250 0 50 100 150 200 250 ◦ min samples = 200.
Illumination level Illumination level
Image
Interpretation of the visual result is as follows. The K-
Means algorithm completely misses the actual clusters of the
road and segments them only by y-value (Fig. 8b).
The result of DBSCAN is actually not so poor, since it
clustered almost all the lanes into separate groups, with the
exception of two lanes, which are located extremely right and
left (Fig. 8c).
HDBSCAN showed better results than other algorithms,
Fig. 4. Example of histogram normalization and its approximation with finding all the lanes on the image, although he found only
Gaussian distribution with rectangle representing calculated threshold part of the lane on the right (Fig. 8d). It is clear that further
Fig. 5. Results of the proposed algorithm in comparison with Otsu binarization

II and I.
The results showed that the second worst result in terms
(a) Initial cropped road image of performance was shown by HDBSCAN with 12.5 FPS
(frames per second). However, it is worth noting that the
implementations of the algorithms are taken from different
libraries, and because of this, the comparison is not very fair.
(b) Road image after binarization
By using downscaled image there is a significant trade-off
between time and precision, which is going to be shown in
following section.

Mean time on Mean time on Mean time on
Algorithm
5th image, sec 15th image, sec 25th image, sec
HDBSCAN 0.7578 0.7517 0.7274
(d) Final result of Clustering
DBSCAN 0.1641 0.1400 0.1438
Fig. 6. Pipeline of clustering module
K-Means 0.2810 0.2925 0.2740

we can fit this points with the spline or parabola and get the TABLE II. P ERFORMANCE COMPARISON OF ALGORITHMS . S CALED
actual lane. BY FACTOR 0.3

Implementations of algorithms in the Python language were Algorithm
Mean time on Mean time on Mean time on
5th image, sec 15th image, sec 25th image, sec
chosen. For HDBSCAN we used native author’s implemen-
tations, DBSCAN and K-Means were used from scikit-learn HDBSCAN 0.0842 0.0783 0.0720
library [23]. DBSCAN 0.0096 0.0094 0.0088
K-Means 0.1202 0.1208 0.1210
The bechmark was conducted on Raspberry Pi 4 model B
with 64-bit ARM Cortex A72 CPU @ 1.5GHz and 4 GB of
RAM. The procedure of the experiment was the following. The B. Experiments
video with road was loaded and every clustering algorithm was
consecutively run on each frame. Frame had the shape (120, In order to test mesurable accuracy the proper dataset for
1200, 3) — height of 120 px, width of 1200 px and 3 RGB testing should have be chosen. We have chosen Unsupervised
colors. The full results of the comparison can be found in Table Llamas (The unsupervised labeled lane markers dataset) [24].
TABLE III. E XPERIMENT RESULTS
V. C ONCLUSION
Scale factor AUC Precision Recall Threshold
In this article we have proposed a novel approach for
1.0 0.36934 0.49383 0.44874 0.42679
detecting lanes on roads that is used for dim and low-light
0.3 0.36649 0.48310 0.33036 0.43664 scenarios. Proposed method was successfully tested on the
real road images taken from highway. The proposed method
of clustering binarized images gave better results than others.
The dataset consists of over 100 000 annotated images, so we As a topic for the future researches we would like to study
took only 536 images from various parts of the dataset in order possibility of using methods of illumination correction and
to test the proposed algorithm. histogram processing for more accurate calculating threshold
value.
Overall, we achieve presion of 49% and AUC of 36%.
As can be seen from Fig. 8 proposed solution is sometimes
segments white car as road(c) and some artifacts are still
present(d). Nonetheless, most of the images are good results
such as (a) or (b). Decrease in precision and AUC can be
explained by two things:

1) Firstly, we must specify ROI(Region Of Interest) on
our own and farther parts of the road are removed
from ROI and therefore are not being detected, which
is okay, because there is no need to detect road mark- (a) Good result. Scale 0.3
ings from 200 meters from a car. But this markings
are marked in the dataset and so in testing phase they
are marked as non-detected.
2) Secondly, algorithm is not finding lanes that are on
the sides, because they are harder to detect and have
low-color intensity.

Future work can be focused on eliminating bad results such
as (c) or (d) in Fig.8 by designing post-processing filter steps
such as curv-fitting and discarding curves, that are inclined
more horizontally, than vertically. Especially this step will be (b) Good result. Scale 1
effective in eliminating noise that is coming from white car
detection, because curve points are grouped in a form of a car,
hence grouped more horizontally.
In order to understand how downscaling images beforehand
affects the precision two tests were conducted. Test results
with original size image and downscaled by factor of 0.3 is
presented in III. As we can see, the decline of precision is not
linearly proportional to decline of image size and this can be
used to process images.
(c) Bad result. Car detection. Scale 0.3

(a) Initial cropped road image

(b) Result of K-Means

(d) Bad result. Road artifacts. Scale 0.3
(c) Result of DBSCAN
Fig. 8. Examples of processed images from the dataset

(d) Result of HDBSCAN
Fig. 7. Comparison between most popular clustering methods
R EFERENCES Robotics in the Makers Era, D. Alimisis, M. Moro, and E. Menegatti,
Eds. Cham: Springer International Publishing, 2017, pp. 104–121.
[1] P. Coppola and F. Silvestri, 1 - Autonomous vehicles and future
[21] L. McInnes, J. Healy, and S. Astels, “hdbscan: Hierarchical density
mobility solutions, P. Coppola and D. Esztergár-Kiss, Eds. Elsevier,
based clustering,” The Journal of Open Source Software, vol. 2, 03
2019. [Online]. Available: http://www.sciencedirect.com/science/article/
2017.
pii/B9780128176962000019
[22] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based
[2] Y. Tian, K. Pei, S. Jana, and B. Ray, “Deeptest: Automated testing
algorithm for discovering clusters in large spatial databases with noise,”
of deep-neural-network-driven autonomous cars,” in Proceedings of the
in KDD’96: Proceedings of the Second International Conference on
40th international conference on software engineering, 2018, pp. 303–
Knowledge Discovery and Data Mining. AAAI Press, 1996, pp. 226–
314.
231.
[3] H. Shin, D. Kim, Y. Kwon, and Y. Kim, “Illusion and dazzle: Adversar-
[23] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
ial optical channel exploits against lidars for automotive applications,”
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander-
in International Conference on Cryptographic Hardware and Embedded
plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch-
Systems. Springer, 2017, pp. 445–467.
esnay, “Scikit-learn: Machine learning in Python,” Journal of Machine
[4] D. Feng, L. Rosenbaum, and K. Dietmayer, “Towards safe autonomous Learning Research, vol. 12, pp. 2825–2830, 2011.
driving: Capture uncertainty in the deep neural network for lidar 3d [24] K. Behrendt and R. Soussan, “Unsupervised labeled lane marker dataset
vehicle detection,” in 2018 21st International Conference on Intelligent generation using maps,” in Proceedings of the IEEE International
Transportation Systems (ITSC). IEEE, 2018, pp. 3266–3273. Conference on Computer Vision, 2019.
[5] H. A. Shiddieqy, F. I. Hariadi, and T. Adiono, “Implementation of deep-
learning based image classification on single board computer,” in 2017
International Symposium on Electronics and Smart Devices (ISESD).
IEEE, 2017, pp. 133–137.
[6] P. A. Lazar and V. Shyam, “Agile development of automated driving
system: A study on process and technology,” Master’s thesis, Chalmers
University of Technology, 2017.
[7] Q. Zou, H. Jiang, Q. Dai, Y. Yue, L. Chen, and Q. Wang, “Robust lane
detection from continuous driving scenes using deep neural networks,”
IEEE Transactions on Vehicular Technology, vol. 69, no. 1, pp. 41–54,
2020.
[8] M. M. Yusuf, T. Karim, and A. F. M. S. Saif, “A robust method for
lane detection under adverse weather and illumination conditions using
convolutional neural network,” in Proceedings of the International
Conference on Computing Advancements, ser. ICCA 2020. New York,
NY, USA: Association for Computing Machinery, 2020. [Online].
Available: https://doi.org/10.1145/3377049.3377105
[9] K.-B. Kim and D. H. Song, “Real time road lane detection with ransac
and hsv color transformation,” J. Inform. and Commun. Convergence
Engineering, vol. 15, 2017.
[10] J. Kim, S. Kim, S. Lee, T. Lee, and J. Lim, “Lane recognition
algorithm using lane shape and color features for vehicle black box,”
in 2018 International Conference on Electronics, Information, and
Communication (ICEIC), Jan 2018, pp. 1–2.
[11] M. S. Drew, J. Wei, and Z.-N. Li, “Illumination–invariant image
retrieval and video segmentation,” Pattern Recognition, vol. 32, no. 8,
pp. 1369 – 1388, 1999. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S003132039800168X
[12] A. A. Assidiq, O. O. Khalifa, M. R. Islam, and S. Khan, “Real time lane
detection for autonomous vehicles,” in 2008 International Conference
on Computer and Communication Engineering, May 2008, pp. 82–88.
[13] J. Wang, W. Hong, and L. Gong, “Lane detection algorithm based on
density clustering and ransac,” in 2018 Chinese Control And Decision
Conference (CCDC), June 2018, pp. 919–924.
[14] N. Otsu, “A threshold selection method from gray-level histograms,”
IEEE Transactions on Systems, Man, and Cybernetics, vol. 9, no. 1,
pp. 62–66, 1979.
[15] R. N. Hota, S. Syed, S. Bandyopadhyay, and P. R. Krishna, “A simple
and efficient lane detection using clustering and weighted regression,”
in Proceedings of the 15th International Conference on Management of
Data, 2009.
[16] S. Eswar, “Noise reduction and image smoothing using gaussian blur.”
Ph.D. dissertation, California State University, Northridge, 2015.
[17] G. Sharma and R. Bala, Eds., Digital Color Imaging Handbook, ser.
Electrical Engineering & Applied Signal Processing Series. CRC Press,
2017.
[18] E. H. Land, “The retinex theory of color vision,” Scientific american,
vol. 237, no. 6, pp. 108–129, 1977.
[19] A. Bovik, The Essential Guide to Image Processing. Elsevier Science,
2009.
[20] J. Tani, L. Paull, M. T. Zuber, D. Rus, J. How, J. Leonard, and A. Censi,
“Duckietown: An innovative way to teach autonomy,” in Educational