=Paper=
{{Paper
|id=Vol-2267/611-614-paper-117
|storemode=property
|title=Convolutional neural networks for self-driving cars on GPU
|pdfUrl=https://ceur-ws.org/Vol-2267/611-614-paper-117.pdf
|volume=Vol-2267
|authors=Boris V. Tiulkin,Nataliia V. Kulabukhova
}}
==Convolutional neural networks for self-driving cars on GPU==
https://ceur-ws.org/Vol-2267/611-614-paper-117.pdf
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
CONVOLUTIONAL NEURAL NETWORKS FOR SELF-
DRIVING CARS ON GPU
B.V. Tiulkin a, N.V. Kulabukhova b
Faculty of Applied Mathematics and Control Processes, Saint Petersburg State University, 13B
Universitetskaya Emb., St Petersburg 199034, Russia
E-mail: a st050135@student.spbu.ru, b n.kulabukhova@spbu.ru
Self-driving vehicles are considered to be safer than those driven by humans. Since they are always
aware of what is happening around them and focuses on all the details. But to be really safe and
respond to all the events happening around the drones need to process information and make decisions
in the shortest possible time. The challenge is to teach how to drive a vehicle without human with the
help of deep learning power using visual data from the cameras installed on the machine. The problem
is to process the amount of data in the real time. Convolutional neural networks (CNNs) are used for
training data. And the idea of how to use CNNs on graphical processing units is described.
Keywords: Self-Driving Cars, High-Performance Computing, Convolutional Neural Networks, GPU
© 2018 Boris V. Tiulkin, Nataliia V. Kulabukhova
611
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
1. Introduction
Currently, there are a large number of projects for the development of unmanned vehicles. The
most famous such as Mobileye by Intel [1], WayMo by Google [2], or even self-driving taxi from
Yandex Research [3]. The basic principle of such systems is the processing of data collected from
cameras and lidars. In the case of Intel, any sensors are absent, since only visual data obtained from
cameras on the machine is analyzed. However, processing of a large number of images in real time is
required.
Main problems for practically every self-driving project:
1. speed of data processing, because it is necessary to obtain them in real time,
2. driving in the absence of road markings,
3. influence of weather conditions and lighting,
4. motion obscure visual landmarks,
5. actions in unforeseen situations.
Convolutional Neural Networks has been used for commercial purposes for more than twenty
years. [4]. But two important events have served as a great impetus to the popularity of these neural
networks in recent years. First, large, labeled datasets such as the ImageNet Large Scale Visual
Recognition Challenge (ILSVRC) [5] are now widely available for training and validation. Another is
that CNNs learning algorithms can now be implemented on massively parallel graphics processing
units (GPUs), greatly speeding up learning and output.
2. Initial data
A person drives a car. It has cameras for data collection. The result is a set of frames, for each
there is a value of the angle of rotation at that time.
Data is collected under different weather conditions, lighting and time of day. They are sorted
on the basis of the above conditions.
To train a network to get out of a bad situation, we expand dataset by adding artificial
perturbations.
Figure 1. Diagram of a one layer convolutional network: convolution followed by downsampling
3. Selection of neural networks
We use convolutional neural networks because they, by reason of their architecture, are well
suited for solving complex graphic image recognition and classification tasks [6].
More specifically, the choice was due to the following advantages:
‒ Algorithms using convolutional neural networks are invariant to various distortions,
such as camera rotation, uneven the distribution of light in the image, horizontal or
vertical shifts and other.
612
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
‒ CNN does not require the allocation of large amounts of memory for storing the
extracted features in a work process.
‒ Fast learning speed, which is achieved by reducing the number of parameters used.
The performance of a CNN is several times greater than the performance of other neural
networks used in recognition tasks.
The idea of convolutional neural networks is the alternation of convolutional layers and layers
of pooling (subsample, subsampling) (see figure 1). The meaning of the convolution operation is that
each fragment of the image is multiplied by the matrix (core) of the convolution element by element,
and the result is summed and recorded in the similar position of the output image [7].
Convolutional neural networks are particularly effective for image recognition tasks, because
the convolution operation captures the 2D nature of images.
4. Neural networks on GPU
To solve the problem requires processing a large amount of data in real time. In order to speed
up the process, the calculations will be performed on the GPU.
Comparing with CPUs, GPUs are better suited for machine learning, because technical
features help them perform at the same time a large amount of data, which is used to train the neural
network presented in a matrix form. But there are some difficulties associated with a limited volume
of the GPU memory. Therefore, it is better to use heterogeneous environment [8] for such kind of
problems, where the graphical accelerators are used for calculations during training and testing data,
and the analysis is made on CPU. The schematic view of the working process is shown on figure 2.
Figure 2. Schematic view of the system
5. Simulation
Lidars, cameras and radars scan everything around the car. The quality of information-pictures
and parameters from sensors - allows the car to analyze road markings, road signs, vehicles, people,
animals and all other elements of the visual environment. Thus, the ability to predict the further course
of moving objects, the vehicle determines what to do next. But to do this, the data must be worked out.
Since, for example, when we use a training set of data obtained during the control of the car by the
613
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
driver, there is a problem that a person is naturally not able to constantly drive the car at the same
distance from the roadside. There are a number of other things to consider before using the data.
Since the classification can take countless hours, it is usually used a larger image of the
objects. Thus, the database already contains two-dimensional preprocessed contours of the required
objects. Such as road signs, people, bicycles, other vehicles, trains and traffic lights.
In addition, generally processed images already contain notes for vehicles in cases where
driving decisions are most difficult, such as at busy intersections, cluttered road systems or in the
presence of multiple lanes.
For the described architecture of convolutional neural network was chosen Python 3.5
language. This is due to the fact that the language is well developed in the machine learning and has a
large number of libraries. Was selected the library TensorFlow [9], as it allows computation as on the
GPU.
The challenge is to select the weights while minimizing the root-mean-square error between
the steering output over the network and the driver data. For training, the method of back propagation
of the error is used.
6. Brief conclusions and development prospects
The paper deals with the topic of learning to drive unmanned vehicles, the method of its
realization by means of convolutional networks is chosen. Hereinafter to test the effectiveness of this
method, a practical implementation.
We test our neural network for speed and accuracy. To improve the work of the CNN, increase
its stability and prevent overfitting, try the dropout method (subnet training method with throwing out
random single neurons).
Next, we try various variable network parameters, such as the number of layers, the core
dimension of the bundle for each of the layers, the number of cores for each of the layers, the step of
the core shift when processing the layer, the need for sub-sampling layers, the degree of dimension
reduction, the transfer function neurons, the presence and parameters of the output fully connected
neural network at the output of the convolutional. These parameters greatly affect the operation of the
neural network and they are selected only empirically.
We test on a different amount of data with artificial distortions.
Testing with different parameters and changing the neural network, we will achieve
the best result.
References
[1] Amnon Shashua, If You Can Drive in Jerusalem You Can Drive (Almost), Anywhere,
https://newsroom.intel.com/editorials/if-you-can-drive-jerusalem-you-can-drive-almost-anywhere/
[2] WayMo Technology, https://waymo.com/tech/
[3] Yandex.Taxi Solution, https://yandex.ru/promo/taxi/sdc
[4] L. D. Jackel, D. Sharman, Stenard C. E., Strom B. I., , and D Zuckert. Optical character
recognition for self-service banking. AT&T Technical Journal, 74(1):16–24, 1995
[5] Large scale visual recognition challenge (ILSVRC). URL: http://www.image-net.org/
challenges/LSVRC/
[6] Bojarski M. et al. End to end learning for self-driving cars //arXiv preprint
arXiv:1604.07316. 2016
[7] Haykin S. Neural networks: a full course, 2nd edition. - Williams Publishing House, 2008
(in Russian)
[8] John Levesque, Aaron Vose, Programming for Hybrid Multi/Manycore MPP Systems, CRC Press
Taylor & Francis Group, p.374, 2018
[9] TensorFlow API Documentation https://www.tensorflow.org/api_docs/
614