International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 18 (2018) pp. 13945-13949
© Research India Publications. http://www.ripublication.com
13945
Object Recognition in Dynamic Environments with Convolutional Neural
Networks
Robinson Jiménez-Moreno
1
, Oscar Avilés Sánchez
2
, Diana Marcela Ovalle
3
1
Professor, Department of Mechatronics Engineering, Nueva Granada Military University, Bogotá, Colombia.
2
Professor, Department of Mechatronics Engineering, Nueva Granada Military University, Bogotá, Colombia.
3
Professor, Department of Electronics Engineering, Distrital Frco. Jose de Caldas University, Bogotá, Colombia.
Abstract
Deep learning techniques have revolutionized pattern
recognition systems over the last decade. Within these
techniques, convolutional neural networks (CNN) have
demonstrated a high performance in the recognition of objects
in images. This article presents the evaluation of a CNN with
93% accuracy in the recognition of objects of four categories,
trained with images at a fixed distance and the variations of said
performance when evaluating said network with images of the
trained categories, with variations of distance, evidencing a
degradation of said performance by 30%. This makes it
possible to establish the need for a structured convolutional
neural architecture that, in function of the advantages of
training of fixed distance by parallel layers, reinforces learning
to accuracy values higher than 90%, whose incidence is
demarcated in robotic mobile applications where the camera
approaches or distances from the object.
Keywords. Machine vision, convolutional neural network,
object recognition, MATLAB, Image RGB-D.
INTRODUCTION
Machine vision algorithms were usually developed in three
basic stages, the one of pre-processing of image for adequacy
of the same in size, elimination of noise and others, stage of
image processing oriented to extraction of characteristics and a
stage of recognition of patterns or classification, the latter
usually developed through neural networks. With the
development of different deep learning techniques [1], and
within these, of convolutional neural networks (CNN) [2],
pattern recognition systems have become more versatile. A
clear example of this, CNNs are now widely used in pattern
recognition [3] and image identification applications [4], with
very high accuracy ranges above 90% in most cases, and with
the ability to integrate the traditional steps described at the
beginning.
Although they are still developing techniques [5], CNN offer
solutions in different areas [6], such as artificial intelligence,
robotics, medicine, intelligent control systems and the
development of intelligent city schemes such as traffic control
applications, etc. So that, as they are applied in different
environments, opportunities for developing these techniques
are found, as a function of the capabilities they offer.
In the field of robotics, several developments have been
presented that allow the interaction of a robot with its medium,
for it is necessary to capture and interpret it, which is achieved
by image capture sensors such as cameras and RGB-D sensors,
i.e., that apart from the image, they give the depth to which the
objects in the captured scene are. CCNs have already been
involved in such applications [7] [8] at the level of conventional
RGB and depth [9-11]. However, mobile robotics applications
are still under development and, as mentioned in [12] [13], the
combination of information of RGB-D and CNN presents an
important nucleus for the development of applications of this
type.
When a robot moves, the distance from the camera to the object
varies, which presents a dynamic learning characteristic of the
object. It is clear that an object presents different perspectives
depending on the distance in which it is observed, as the focus
of vision moves away from an object, specific features of the
object are lost and the outline is the most relevant information,
as it approaches the object specific details appear. This can
present a challenge in the performance of a CNN when
identifying an object, when it is used from a mobile robot that
approaches or distances an object of interest. In the state of the
art this analysis is not found, for which it is addressed in the
present article.
This article is organized as follows. In section 2 the CNN
architecture to be used is presented, the generalities of the
training of this type of networks and the results of the training
for the recognition of objects at a fixed distance. In section 3 it
is presented the analysis of the trained network with distance
variations in the test database and a new training including this
new database. In section 4 the conclusions of the evaluation of
the network in dynamic atmospheres of image capture input for
prediction of CNN are presented.
Convolutional neuronal network architecture
Robotic agents handle different path planning schemes to
achieve a specific goal. For this task it is fundamental to
recognize the environment in which it moves and the objective
that seeks, typically this objective is an object, as is the case of