UDC 629.113.006
Firov N.A., Kobersi
I.S.
The scientific adviser: PHD, prof. Finaev V.I.
Development
of algorithm for identifying stationary ground obstacles using neural network in the tasks of vehicle parking
The article discusses development of algorithm for identifying stationary ground obstacles using neural network in the solve tasks of vehicle parking. Formulated requirements to this algorithm identify obstacles. Proposed the neural network structure identification, determined the quantity of neurons in the input,
hidden and output layers. And in this paper was develop optimizing genetic algorithm of neural network parameters.
Currently, the
systems bound to application of the theory of detection of objects in real time
under various conditions of driving of transportation facilities, in particular
parking, are very relevant. The need in using such systems due to the fact that
the most frequent occurrence of accidents occur in the case of parking area,
when driver visibility is very limited.
Algorithm of identification is intended to provide timely
information on the location, size, and distance to obstacles, and to warn the driver of danger, if their distance is small too.
To solve the above problems, the
most optimal solution is to apply the
theory of modeling systems [2], the theory of artificial intelligence [1, 3], and the basis of the theory of
identification algorithms and automation [1].
The article
discusses the rationale and relevance of the problem, proposed parameters which will identify the obstacles, an overview of the existing methods for identifying the obstacles relative to the vehicle and carried out the requirements to the technical means of identification obstacles.
Identification of obstacles implements by measuring two parameters (Fig. 1) : α – angle offset from the normal vector X; R – maximum distance (radius) to a total reflection or scattering signal emitted by means of detection, in which the object can be reliably detected;
Fig. 1. Detection
of various obstacles when changing parameters TC.
Quantity of sensors used should be selected
so that as to eliminate "deadbands",
that is, the sensors should detect obstacles within the space transport vehicle movement at a parking with a maximum radius
of obstacle detection to provide the necessary control for the
future (Fig. 2).
Fig. 2. Zones detections by sensors.
Based on the Fig. 2, also should be noted that the scope of each sensor
intersects with adjacent areas of neighboring sensors, resulting in the measurement of
parameters appear identical values affecting the accuracy and performance of the system.
The same angle of the intersected one identified argument obstacles can be defined as
follows:
(1)
where ,
– angle when measuring in two adjacent areas of visibility sensors.
For the solution of a researched problematic, it is proposed to use neural network identification, as
an identifier of a self-learning networks use genetic algorithms, the
structure of the proposed model is shown in Fig. 3:
Fig. 3. The model
structure system of identification with genetic algorithm of optimization.
Number of elements of the input layer is defined amount of input
data transmitted from all sensors at a time
(2)
On an input each neuron receives data arrays in the form of
, (3)
where i – the ordinal number of the input neurons in the
network, j – index - value of the vector of input variables in the sector of the security zone.В результате
значения с входного слоя поступают на вход скрытого слоя нейронной сети.
(4)
The signals from the hidden layer neurons arrive at the output layer, which form a network response. In this case, the number of neurons in the output (last) layer is one. This layer displays the
results in the form of a one-dimensional matrix.
W = [
…
], (5)
which consists of parameters describing the position obstacles relatively transport vehicle and its shape.
The structure of a designed neural network system of
identification is shown in Fig. 4.
Fig. 4. The structure of a designed neural network system of
identification.
In this case optimization of parameters of the neural
network is a major part of its construction, the function of which is to correct its parameters to
achieve accurate results. As an optimization algorithm proposed a genetic algorithm that can traverse the same value of three adjacent sectors of visibility detection by the following sequence:
Step 1. Generation
of the initial population in the form of:
,
where each line represents the values defined for the i-th neuron of the input
layer;
- angle vector and distance i-th neuron j-th sensor.
Step 2. Development of the objective function.
The objective function is a function of the intersection
of repeatable values three
overlapping sectors. In general form objective function can be represented as:
(6)
where =
,
=
,
=
Step 3. The introduction
of stop condition of the algorithm.
(7)
Rule 1: If at least one parameter it
is necessary to leave one repeatable value and pass it to the output-dimensional
array.
Rule 2: If , then transfer all the values in the output-dimensional
array.
Step 4. Output of results and the transition
to the next elements of arrays of input variables.
The structure of the genetic algorithm optimization neural network identification obstacles is shown in Fig. 5.
Process of
transformation and remove repeatable values three
overlapping sectors of visibility sectors for correct identification and
reduce the output time solution is the justification
for the lack of operators of the genetic algorithm,
because in them there is no need, and they do not perform their functions.
The genetic algorithm performs
the task of training a neural network to
eliminate repeatable values of
the measured data and the transformation of the
parallel input of data in serial
output array using a neural network.
Fig. 5.
The structure of the genetic algorithm optimization neural network identification obstacles
Conclusions:
In the article proposed a structure of
the neural network identification, defined
number of neurons in the input, hidden and output layers. Developed objective
function of genetic algorithm that
performs task of removing repeatable
values intersecting zones of visibility sensors and
transformation parallel input of
data in serial output array, which
reduces the time-to-result of
identification, by reducing the data to be processed, which is important in the future to ensure the required system
performance as a whole.
References
1. Коберси И.С., Белоглазов Д.А. Интеллектуальная адаптивная гибридная обучаемая система управления транспортными средствами. Известия ЮФУ. Технические науки. Тематический выпуск «Методы и средства адаптивного управления в электроэнергетике». – Таганрог: Изд-во ТТИ ЮФУ, 2010, №1 (102). – 260 с., С. 110-117.
2. Финаев В.И. Моделирование при проектировании информационно-управляющих систем: Учебное пособие. - Таганpог: Изд-во ТРТУ, 2002.
3. Коберси И.С., Шадрина В.В. Применение нейронных сетей для управления энергопотреблением. Известия ЮФУ. Технические науки. Тематический выпуск. «Актуальные проблемы производства и потребления электроэнергии». – Таганрог: Изд-во ТТИ ЮФУ, 2008. № 7 (84) – 240 c., С. 190-1964.