Вы находитесь на странице: 1из 4

Machine Learning: Super Vector Machine for traffic

signs recognition
Rosine Dias, Emilien Westelynck
Polytechnic Institute of Advanced Science (IPSA), Paris
rosine.dias@ipsa.fr, emilien.westelynck@ipsa.fr

examples with each their own slack variables (we will


explain how these work), each marked as belonging to either
Abstract – In this paper we display what we learned about the classification or regression analysis. SVM can solves many
concepts of machine learning and artificial intelligence, the practical problems, such as small sample, non-linear, high
different methods to implement them with different degrees of
freedom and cleverness. We chose to present an industrial project dimension number and local minimum points and has been
of this state-of-the-art technology: an application of support vector applied in pattern classification, regression forecasting,
machine to autonomous vehicle. This architecture permits a clever probability estimation, control theory and so on [3][4].
training of the system: the vehicle identifies what it sees on the However, most common use of SVM is in classification
road and classifies the object. This is done to prioritize the security problems, thus the recognition in real time of traffic sign on the
of the passengers without forgetting the environment outside.
road.

Machine Learning, Support Vector Machine, Autonomous The rest of the paper is organised as follows: Section II
vehicle, Traffic signs describes Support Vector Machine, how it works and the basic
performance. Section III shows the application on traffic signs
I. INTRODUCTION recognition on cars, its use in adaptive cruise control and
The number of data science & machine learning projects is performance comparison. Conclusion of this project is given in
growing exponentially since the last decade. We use the quick Section IV along with future directions.
and clever thinking artificial intelligence (AI) to complete a
variety of tasks: anti-spam mailbox, image sorting or self- II. SUPPORT VECTOR MACHINE METHODS
driving cars. The latter represents the future of transports, A. Method description
which is why most multinational companies invest billions in
research and development, such as Ford, which plans to invest As explained above, Support Vector Machine is used to solve
nearly $4 billion by 2023 [1]. Autonomous vehicles are capable classification problems. To use this machine learning method,
of self-driving from point A to point B, without human we plot each data set items as a point in n-dimensional space
intervention. They recognize their environment with their (where n is number of features that you know in advance) with
numerous cameras, infrared or ultrasonic sensors and can adapt the value of each feature being the value of a particular
their speed accordingly with adaptive cruise control (ACC), coordinate. Then, we perform the so-called classification by
performing avoidance manoeuvres if a hazardous situation is finding the hyperplane (basically a line in space) that
detected. It is in this perspective that Support Vector Machine differentiate the classes the most efficiently possible.
is used for Vehicle detection, ACC and tracking [2].

The classic definition of Machine Learning (ML) is as


follows:

“A computer program is said to learn from experience with


respect to some class of task and performance measure, if its
performance at that task -as measured by the performance-
improves with experience.”

We can classify ML in two broad categories:

- Supervised Learning
- Unsupervised Learning

Support Vector Machine is part of the Supervised Learning


category meaning that the system is given a set of training Figure 1: Classification of two different data sets
Figure1. Above shows a simplified view of the classification III. APPLICATION ON TRAFFIC SIGNS RECOGNITION
of two data set performed by Support Vector Machine. This
method follows two very important rules: A. Adaptative cruise control (ACC)
- Separate correctly the data sets
- Maximize the distance between the two decisions 1. Explanation
boundaries or support vector (dotted lines in Figure1.)
Adaptive Cruise Control (ACC) allows the vehicle to adapt
Mathematically, we use the following notation: to the speed of the traffic regulation by detecting the traffic sign.
The car is capable to analyze the four different signs (warning
- 𝑥 : a feature vector (the input of the SVM). 𝑥 ∈ 𝑅𝑛 , signs, prohibition, obligation and informative sign) depending
where n is the dimension of the vector; on their form, their colors and on their textual differences. The
- 𝑦: the class (the output of the SVM). 𝑦 ∈ {−1,1}, so that main difficulties for the car are to determine which traffic sign
the classification task is binary; goes to which class because of the human or natural
- 𝑤, 𝑏: the parameters of the SVM: we need to learn them deterioration (we will introduce a solution train the SVM to
using the training set; read damages traffic sign in 2 parts).
- 𝑁: the number of sets; To cope to this situation ACC vehicle are composed on
- (𝑥𝑖 , 𝑦𝑖 ): ith sample in the dataset. (∀𝑖 ∈ {1, … , 𝑁}) ultrasonic sensor. These sensors positioning in front of the car
are capable to analyze if the vehicle in front are slowing or if
Then we can rewrite the two rules: there’s an obstacle. Taking the example of a car in front, if the
sensor detects the vehicle she will automatically slow and
calculate the time gap between the two, if the sensor detects that
the vehicle is no longer in front, it will accelerate to adapt to the
speed limit.

(1) 2. Example of application


The problem is to optimize the two rules, thus optimize the
quadratic optimization problem given in Eq (1) [7] Actually, Adaptive Cruise Control are already existing as you
can see with Volkswagen. This car is maintaining a speed set
by the driver and also adapts to the speed of the traffic
B. Performance: Pros and cons regulation. So, the car takes into account the limit of speed by
analyzing traffic sign with an embedded camera and a
On one hand, one of the very fundamental advantage of SVM navigation system.
is the classification of data sets when the margin of separation
is clear: in other term, when the two samples don’t have any B. Performance
similarities and no outlier (sample mixing with data set that
isn’t its own) the hyperplane can be determined very efficiently. 1. Explanation of the study
Moreover, the more difficult is the training sample, the better
the method will cope with simple samples improving the Several applications were made to measure the performance
processing time. of SVM for traffic sign recognition. To demonstrate the
Other assets of SVM: it’s very effective dealing with high performance of these sign recognition we will focus on multi-
dimension spaces, plus is effective when the number of class classification problem [4].
dimensions is greater than the number of samples. Traffic signs in Colombia are classified into vertical and
horizontal signs. The objective of the application is to classify
On the other hand, it doesn't generally perform well, mostly the vertical sign in a video stream. In order to achieve this test
when we have large data sets increasing the training time. The a data set was introduced containing about one-third of images
cost of this method is high, it’s calculated using the following with a given traffic signs, called positive examples and two
formula: third of images containing other traffic signs, called negative
examples.
The first point of the study was to modify the original traffic-
sign provided by the Ministry of Transport to incorporate some
geometric transformation (pitch, roll and yaw) and variations in
(2) terms of color (saturation) and illuminations. The aim of this
modification was to train the SVM to recognize the traffic sign
in all the configuration that can be possible on the road (i.e.
illumination source, vandalization) and to improve the
robustness of the 47 classifiers by exposing them in challenging
situations. After the modification 3872 positives images were
created for each class.

Figure 2 : Each case represents one class

The training data set was composed of a feature matrix x and


the target vector y that will indicate if the example belongs to a
class or not. The matrix is composed as follows, with m the
number of examples, n the number of features and 𝑦 ∈ {0,1}.

Figure 5 : Confusion matrix and formula for each class

2. Related study

Another study [8] used also SVMs to conduct a pattern


recognition and traffic sign detection similar to the
Universidad de San Buenaventura.
Their technic was more focused on the traffic sign
detection with a simplified shape/color classification using
Figure 3: Feature matrix and target vector image sequences. They implemented the following
illustrated method:
For each class they divided the available data in three
different data sets, one data set will be used for training (60%
of the data that represents 7022 examples), one data set for
model selection (20% of the data that represents 2326 examples)
and the last and third data set to test the system with the same
number of examples that the second one. Figure 6: Block diagram of TS detection and recognition
The second point of the study was to configure the SVM.
Each SVM was trained using the Sequential Minimal They obtained the following results:
Optimization. To configure the second data set they used the
SVM kernel with The Gaussian Radial Basis Function (RBF)
because of his best performance.
To measure the performance of the system they introduce a
confusion matrix for binary classifier. This confusion matrix
will be used when the output of the binary classifier is
discrete. The performance will be calculated based on the
number of true positives TP, false positive FP, false negative
FN and true negative TN. Figure 4 : Pattern recognition results using edge related POI as
feature vector

In this table the true positive represents percentage of number


of traffic sign images classified correctly, the false positive
indicates the percentage of images classified wrongly and the From these two tests we can deduct that the second got the
false negative the percentage of missed classifications. highest performance ratio. However, the two tests can’t be
Their results are less performant than the main study that we directly compared because of the composition of the two data
worked on, but it uses less training samples. sets. Indeed, the number of positive examples was not the same
for each classifier in the two tests. In the first test each classifier
was composed with 33.32% of positive examples and in the
3. Results second test each classifier has approximately 2.13% of positive
examples.
The aim of the first test was to experiment individual binary
classifier with the same number of training examples as before. IV. CONCLUSIONS
The result obtained for the 47 classifiers are shown in the
following graph. From this image we can see that the In this report, we learnt how a machine learning method: the
performance for each classifier is approximately 95%. Support Vector Machine works for classifying different data
sets efficiently. We saw the application of this particular
method in the automobile industry with one of the assets of the
autonomous car: the traffic sign recognition in adaptive cruise
control. One of the future developments we can think of is the
application of this method not only to sign recognition, but also
to any agent that an autonomous vehicle can encounter such as
other drivers, cycles, pedestrians etc. This will allow the vehicle
to be fully “conscious” of its environment, adapting its behavior
to the circumstances.
ACKNOWLEDGMENTS
We would like to thank our professor of Machine Learning,
Joana Maria Frontera Pons, associate professor of IPSA who
Figure 7 : Results obtained with the first test introduced us to this field of study.
REFERENCES
The second test was to evaluate the performance of classifiers
[1] F. Holland, “Here's how Ford's autonomous vehicles will shake up ride
working together. From this test we can highlight two major hailing and Delivery services”, CNBC, Nov 15, 2018
information.
Firstly,302 examples of 36425 examples were no recognize [2] R. Uppala, “Vehicle detection and tracking using a Support Vector Machine
for autonomous vehicle applications”, Medium, Aug 15, 2017
by any classifier and secondly, a traffic sign can’t be recognized
by more than one classifier at the same time. The results [3] K. Yang, L-L. Zhao, “Application of the improved Support Vector Machine
obtained are shown in the following image, the performance on vehicle recognition”, Hebei University of Science and Technology,
China, July 2018
ratio for each classifier is above 99%.
[4] J-A. Gomez, S. Bromberg, “Automated traffic sign recognition system using
computer vision and Support Vector Machines”, Universidad de San
Buenaventura, Cali, Columbia, July 2014

[5] J. Zhang, Y. Liao, S. Wang, J. HanStudy, “Driving Decision-Making


Mechanism of Autonomous Vehicle Based on an Optimized Support
Vector Machine Regression”, Shandong University of Science and
Technology, China, December 2017

[6] K. Lim, Y. Hong, Y. Choi, H. Byun, “Real-time traffic sign recognition


based on a general-purpose GPU and deep-learning”, Beihang
University, China, March 6, 2017

[7] O. Chapelle "Training a support vector machine in the primal." Neural


computation Volume.19, 2007, MIT, USA pp. 1155-1178

[8] C.G. Kiran, V. Lekhesh, V. Abdu Rahiman, K. Rajeev, "Traffic Sign


Figure 8: Results obtained with the 2nd test Detection and Pattern Recognition using Support Vector Machine”,
Thiruvananthapuram, India pp. 87-90, 2009

Вам также может понравиться