U.S. patent application number 17/082225 was filed with the patent office on 2021-06-10 for deep learning-based object recognition system and method using pir sensor.
This patent application is currently assigned to Pusan National University Industry-University Cooperation Foundation. The applicant listed for this patent is Pusan National University Industry-University Cooperation Foundation. Invention is credited to Donghyun KIM, Jongdeok KIM, Jaemin LEE.
Application Number | 20210174112 17/082225 |
Document ID | / |
Family ID | 1000005182870 |
Filed Date | 2021-06-10 |
United States Patent
Application |
20210174112 |
Kind Code |
A1 |
KIM; Jongdeok ; et
al. |
June 10, 2021 |
DEEP LEARNING-BASED OBJECT RECOGNITION SYSTEM AND METHOD USING PIR
SENSOR
Abstract
Disclosed is a deep learning-based object recognition system
which includes a data collection process executor configured to
perform a data collection process of collecting values of the PIR
sensor according to a sampling period; a data classification
process executor configured to perform a data classification
process of inputting the collected values of the PIR sensor to a
model of an artificial neural network and transmitting a result of
the inputting the collected values to a cloud using a RESTful API;
and an object recognition cloud system configured to store
information received from the data classification process executor
in a database, transmit the information when a web application
requests the information, and represent information collected and
classified by devices using the RESTful API.
Inventors: |
KIM; Jongdeok; (Busan,
KR) ; LEE; Jaemin; (Changwon-si, KR) ; KIM;
Donghyun; (Busan, KR) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Pusan National University Industry-University Cooperation
Foundation |
Busan |
|
KR |
|
|
Assignee: |
Pusan National University
Industry-University Cooperation Foundation
Busan
KR
|
Family ID: |
1000005182870 |
Appl. No.: |
17/082225 |
Filed: |
October 28, 2020 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06K 9/6256 20130101;
G06K 9/00771 20130101; G06K 9/6232 20130101; G06K 9/00335 20130101;
G06K 9/40 20130101; G06K 9/00677 20130101; G06K 9/2018
20130101 |
International
Class: |
G06K 9/20 20060101
G06K009/20; G06K 9/00 20060101 G06K009/00; G06K 9/62 20060101
G06K009/62; G06K 9/40 20060101 G06K009/40 |
Foreign Application Data
Date |
Code |
Application Number |
Dec 5, 2019 |
KR |
10-2019-0160873 |
Claims
1. A deep learning-based object recognition system using a passive
infra-red (PIR) sensor, the system comprising: a data collection
process executor configured to perform a data collection process of
collecting values of the PIR sensor according to a sampling period;
a data classification process executor configured to perform a data
classification process of inputting the collected values of the PIR
sensor to a model of an artificial neural network and transmitting
a result of the inputting the collected values to a cloud using a
RESTful API; and an object recognition cloud system configured to
store information received from the data classification process
executor in a database, transmit the information when a web
application requests the information, and represent information
collected and classified by devices using the RESTful API, wherein
the data process executor is further configured to limit two or
more duplicate values so as not to be produced to prevent overload
on a microprocessor due to calculation by the artificial neural
network, a calculation time t.sub.n of the data classification
process is defined as t.sub.n=t.sub.after.sub.n-t.sub.prev.sub.n, a
limitation of a number of the data classification processes
according to time is defined by
t.sub.after.sub.n<t.sub.prev.sub.n+1, and the data collection
process executor, the data classification process executor, and the
object recognition cloud system are each implemented via at least
one processor.
2. The system of claim 1, wherein the PIR sensor provides an output
on the basis of a voltage scale within a specific area to produce
frequencies by performing a fast Fourier transform (FFT) on an
analog signal sample and use the frequencies as a convolutional
neural network (CNN) feature vector.
3. The system of claim 1, wherein the object recognition cloud
system comprises: a RESTful API framework configured to store
information received from the data classification process executor
in the database and transmit the information when the web
application requests the information; and a monitoring web
application configured to represent information collected and
classified by the devices using the RESTful API.
4. The system of claim 1, wherein the data collection process
executor inputs the collected values of the PIR sensor to the model
of the artificial neural network, and the artificial neural network
is trained with signal patterns of motions of a human and an animal
on the basis of a difference in waveform between infrared rays from
a human and an animal to identify whether a recognized object is a
human or an animal and to classify input data as one of three
results, including no object, an animal, and a human, and indicates
a distance to the object.
5. A deep learning-based object recognition method using a passive
infra-red (PIR) sensor, the method comprising: performing a data
collection process of collecting values of the PIR sensor according
to a sampling period; performing a data classification process of
inputting collected values of the PIR sensor to a model of an
artificial neural network and transmitting a result of the
inputting the collected values to a cloud using a RESTful API; and
recognizing an object by storing information received in the data
classification process in a database, transmitting the information
when a web application requests the information, and representing
information collected and classified by devices using the RESTful
API, wherein the performing of the data classification process
comprises limiting two or more duplicate values so as not to be
produced to prevent overload on a microprocessor due to calculation
by the artificial neural network, a calculation time t.sub.n of the
data classification process is defined as
t.sub.n=t.sub.after.sub.n-t.sub.prev.sub.n, and a limitation of a
number of the data classification processes according to time is
defined by t.sub.after.sub.n<t.sub.prev.sub.n+1.
6. The method of claim 5, wherein the performing of the data
collection process comprises inputting the collected values of the
PIR sensor to the model of the artificial neural network and
training the artificial neural network with signal patterns of
motions of a human and an animal on the basis of a difference in
waveform between infrared rays from a human and an animal to
identify whether the recognized object is a human or an animal and
to classify input data as one of three results, including no
object, an animal, and a human, and indicate a distance to the
object.
7. (canceled)
8. The method of claim 5, wherein, when a range of t.sub.n is
i.+-..alpha. ms, the data collection process is performed at time
intervals longer than i.+-..alpha. ms, which is a maximum value of
t.sub.n, to prevent redundant creation of the data collection
process and is created at time intervals of i.+-..alpha. ms.
9. The method of claim 5, wherein the artificial neural network
comprises: a data feature extracting neural network trained with
waveforms of infrared rays from a human and an animal, which are
receivable by the PIR sensor, to extract data features using a
convolutional neural network (CNN) and cope with a change of data
values due to noise and environmental changes; and a feature
correlation identifying neural network configured to identify a
correlation on the basis of several features extracted from one
data sample using a recurrent neural network (RNN) and to derive a
final result.
10. The method of claim 5, wherein the final result is classified
as one of three results, including a human, an animal, and no
object, and represents a distance to the object, wherein the
distance of the final result is continuous data and is obtained
according to a linear regression, the human, the animal, and the no
object each represent a result corresponding to 0 or 1 according to
a presence or absence of an object and are obtained according to
logistic regression, and an algorithm for object recognition is
configured as a combination of the linear regression and the
logistic regression.
11. The method of claim 10, wherein cost functions according to the
linear regression and the logistic regression are different, and a
final cost function is calculated by adding values of cost
functions of the linear regression and the logical regression,
wherein the cost functions of the linear regression and the
logistic regression
Cost.sub.linear=(y.sub.linear-y.sub.linear).sup.2 are expressed by:
Cost.sub.logistic=-y.sub.logistic
log(y.sub.logistic)-(1-y.sub.logistic)log(1-y.sub.logistic), and
the final cost function is calculated by: J ( .theta. ) = 1 m i = 1
m [ Cost linear + Cost logistic ] . ##EQU00003##
Description
ACKNOWLEDGEMENT
[0001] This research was supported by the Korean Government,
Ministry of Science and ICT (MSIT, Republic of Korea), under a
Grand Information Technology Research Center support program
(IITP-2020-2016-0-00318) supervised by an Institute for Information
& Communications Technology Planning & Evaluation
(IITP).
CROSS-REFERENCE TO RELATED APPLICATION
[0002] This application claims priority to Korean Patent
Application No. 10-2019-0160873 (filed on Dec. 5, 2019), which is
hereby incorporated by reference in its entirety.
BACKGROUND
[0003] The present invention relates to an object recognition
system, and more particularly, to a deep learning-based object
recognition system and method using a passive infra-red (PIR)
sensor, in which a human and an object are identified using a new
type of PIR sensor and a machine learning-based object detection
algorithm.
[0004] A passive infra-red (PIR) sensor, which is one of various
sensors capable of detecting objects, outputs time-series data
according to a voltage change due to the pyroelectric effect, and a
value thereof changes as a difference, which changes over time, in
the amount of far-infrared rays emitted in the surroundings is
passively accepted.
[0005] Various factors, which change the amount of emitted
far-infrared rays, include a surface area of an object emitting
far-infrared rays, heat of a surface of the object, and a moving
speed of the object, etc., and the amounts of emitted far-infrared
rays that are changed by the factors are different from each
other.
[0006] In an object detection method of the related art, an object
is detected using a threshold or a simple algorithm. Because a
value of this method is affected by various factors, the
reliability of a determination as to whether an object is present
according to the method of the related art is not likely to be high
and the method is of limited application.
[0007] Machine learning on which many related studies are being
conducted can be utilized for various applications using a PIR
sensor because it is possible to draw conclusions thereby in
consideration of various variables.
[0008] However, when learning is performed to produce acceptable
performance, it requires a high cost to collect a necessary amount
of data and give a meaning to the data.
[0009] Meanwhile, in security systems, numerous light-sensing
machines are being developed to prevent intrusion and generate an
alarm using a PIR-based motion detection sensor.
[0010] However, the PIR-based motion detection sensor that operates
according to a temperature difference between an object and a
surrounding environment is very sensitive when the object moves
closer thereto.
[0011] Therefore, the sensitivity of this sensor is low when the
object is close enough to warm up the surrounding environment, and
the sensor has more problems when an ambient temperature is closer
to a human body temperature, e.g., during summer, than during
winter.
[0012] In addition, the sensitivity of the sensor is likely to be
low when the human body is moving slowly or a cover is present that
blocks heat.
[0013] For example, when a user is holding an umbrella or wearing a
raincoat, the umbrella or raincoat blocks heat from the user's body
and thus it is difficult for the PIR sensor to detect a motion.
[0014] A reduction in the sensitivity of the sensor also occurs
under sunlight.
[0015] An intrusion prevention system using a PIR sensor of the
related art detects a motion on the basis of a threshold of the PIR
sensor. When a digital logic value exceeds a fixed threshold, it is
determined that an intrusion occurred and thus a logic HIGH value
is generated, and otherwise, it is determined that no intrusion has
occurred and thus a logic LOW value is generated.
[0016] The intrusion prevention system using the PIR sensor of the
related art is capable of detecting only whether an object is
present, and it is difficult to identify the detected object.
[0017] Accordingly, there is a need to develop technology for an
object recognition system for identifying a human and an object
using a new type of PIR sensor and a machine learning-based object
detection algorithm.
SUMMARY
[0018] To address the problems of an object recognition system of
the related art, the present invention is directed to providing a
deep learning-based object recognition system and method using a
passive infra-red (PIR) sensor, in which a human and an object are
identified using a new type of PIR sensor and a machine
learning-based object detection algorithm.
[0019] The present invention is directed to providing a deep
learning-based object recognition system and method using a PIR
sensor to increase object recognition performance by extracting
various signals obtained by the PIR sensor and processing PIR data
to extract a frequency component of a signal as a feature
vector.
[0020] The present invention is directed to providing a deep
learning-based object recognition system and method using a PIR
sensor to improve object recognition performance through a learning
method using an artificial convolutional neural network (CNN) and
classification after learning.
[0021] The present invention is directed to providing a deep
learning-based object recognition system and method using a PIR
sensor to increase object recognition performance and usability by
building a real-time-web-based monitoring system using an
object-cloud.
[0022] The present invention is directed to providing a deep
learning-based object recognition system and method using a PIR
sensor to reduce a false-alarm rate using an object recognition
technology of measured data using deep learning.
[0023] Aspects of the present invention are not limited thereto and
other aspects not mentioned here will be clearly understood by
those of ordinary skill in the art from the following
description.
[0024] According to one aspect of the present invention, a deep
learning-based object recognition system using a passive infra-red
(PIR) sensor includes a data collection process executor configured
to perform a data collection process of collecting values of the
PIR sensor according to a sampling period; a data classification
process executor configured to perform a data classification
process of inputting the collected values of the PIR sensor to a
model of an artificial neural network and transmitting a result of
the inputting the collected values to a cloud using a RESTful API;
and an object recognition cloud system configured to store
information received from the data classification process executor
in a database, transmit the information when a web application
requests the information, and represent information collected and
classified by a device using the RESTful API.
[0025] Here, the PIR sensor may provide an output on the basis of a
voltage scale within a specific area to produce frequencies by
performing a fast Fourier transform (FFT) on an analog signal
sample and use the frequencies as convolutional neural network
(CNN) feature vectors.
[0026] The object recognition cloud system may include a RESTful
API framework configured to store information received from the
data classification process executor in the database and transmit
the information when the web application requests the information,
and a monitoring web application configured to represent
information collected and classified by the devices using the
RESTful API.
[0027] The data collection process executor may input the collected
values of the PIR sensor to the model of the artificial neural
network, and the artificial neural network may be trained with
signal patterns of motions of a human and an animal on the basis of
a difference in waveform between infrared rays from a human and an
animal to identify whether a recognized object is a human or an
animal and to classify input data as one of three results,
including no object, an animal, and a human, and indicate a
distance to the object.
[0028] According to another aspect of the present invention, a deep
learning-based object recognition method using a PIR sensor
includes performing a data collection process of collecting values
of the PIR sensor according to a sampling period; performing a data
classification process of inputting collected values of the PIR
sensor to a model of an artificial neural network and transmitting
a result of the inputting the collected values to a cloud using a
RESTful API; and identifying an object by storing information
received from the data classification process in a database,
transmitting the information when a web application requests the
information, and representing information collected and classified
by devices using the RESTful API.
[0029] Here, the performing of the data collection process may
include inputting the collected values of the PIR sensor to the
model of the artificial neural network and training the artificial
neural network with signal patterns of motions of a human and an
animal on the basis of a difference in waveform between infrared
rays from a human and an animal to identify whether a recognized
object is a human or an animal and to classify input data as one of
three results, including no object, an animal, and a human, and
indicate a distance to the object.
[0030] The performing of the data classification process may
include limiting two or more duplicate values so as not to be
produced to prevent overload on a microprocessor due to calculation
by the artificial neural network, wherein a calculation time
t.sub.n of the data classification process may be defined as
t.sub.n=t.sub.after-t.sub.prev.sub.n, and a limitation of the
number of classification processes according to time may be defined
by t.sub.after<t.sub.prev.sub.n+1.
[0031] When a range of t.sub.n is i.+-..alpha. ms, the data
collection process may be performed at time intervals longer than
i.+-..alpha. ms, which is a maximum value of t.sub.n, to prevent
redundant creation of the data collection process and may be
created at time intervals of i.+-..alpha. ms.
[0032] The artificial neural network may include a data feature
extracting neural network trained with waveforms of infrared rays
from a human and an animal, which are receivable by the PIR sensor,
to extract data features using a convolutional neural network (CNN)
and cope with a change of data values due to noise and
environmental changes; and a feature correlation identifying neural
network configured to identify a correlation on the basis of
several features extracted from one data sample using a recurrent
neural network (RNN) and to derive a final result.
[0033] The final result may be classified as one of three results,
including a human, an animal, and no object, and represents a
distance to an object, wherein the distance of the final result may
be continuous data obtained according to a linear regression, the
human, the animal, and the no object may each represent a result
corresponding to 0 or 1 according to a presence or absence of an
object and are obtained according to a logistic regression, and an
algorithm for object recognition may be configured as a combination
of the linear regression and the logistic regression.
[0034] Cost functions according to the linear regression and the
logistic regression may be different, and a final cost function is
calculated by adding values of the cost functions of the linear
regression and the logical regression, wherein cost functions
according to the linear regression and the logistic regression are
expressed by:
Cost.sub.linear=(y.sub.linear-y.sub.linear).sup.2
Cost.sub.logistic=-y.sub.logistic
log(y.sub.logistic)-(1-y.sub.logistic)log(1-y.sub.logistic),
and the final cost function is expressed as:
J ( .theta. ) = 1 m i = 1 m [ Cost linear + Cost logistic ] .
##EQU00001##
BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The above and other objects, features and advantages of the
present invention will become more apparent to those of ordinary
skill in the art by describing exemplary embodiments thereof in
detail with reference to the accompanying drawings, in which:
[0036] FIG. 1 is an overall configuration diagram of a deep
learning-based object recognition system using a passive infra-red
(PIR) sensor according to the present invention;
[0037] FIGS. 2A and 2B are configuration diagrams illustrating
operating characteristics of a deep learning-based object
recognition system using a PIR sensor according to the present
invention;
[0038] FIG. 3 is a detailed configuration diagram of a deep
learning-based object recognition system using a PIR sensor
according to the present invention;
[0039] FIG. 4 is a flowchart of an event processing process of a
deep learning-based object recognition system using a PIR sensor
according to the present invention;
[0040] FIG. 5 is a flowchart of a deep learning-based object
recognition method using a PIR sensor according to the present
invention;
[0041] FIG. 6 is a diagram illustrating a real-time data processing
process; and
[0042] FIGS. 7A and 7B are diagrams for describing a deep
learning-based object recognition algorithm using a PIR sensor
according to the present invention.
DETAILED DESCRIPTION
[0043] Hereinafter, exemplary embodiments of a deep learning-based
object recognition system and method using a passive infra-red
(PIR) sensor according to the present invention will be described
in detail.
[0044] Features and advantages of the deep learning-based object
recognition system and method using the PIR sensor according to the
present invention will become apparent from a detailed description
of embodiments to be described below.
[0045] FIG. 1 is an overall configuration diagram of a deep
learning-based object recognition system using a PIR sensor
according to the present invention. FIGS. 2A and 2B are
configuration diagrams illustrating operating characteristics of a
deep learning-based object recognition system using a PIR sensor
according to the present invention.
[0046] A deep learning-based object recognition system and method
using a PIR sensor according to the present invention are capable
of identifying a human and an object using a new type of PIR sensor
and a machine learning-based object detection algorithm, in which a
frequency component of a signal is extracted as a feature vector by
extracting various signals obtained by the PIR sensor and
processing PIR data, thereby improving object recognition
performance.
[0047] In particular, object recognition performance may be
improved by a learning method using an artificial convolutional
neural network (CNN) and classification after learning, and the
performance and use of object recognition may be improved by
building a real-time web-based monitoring system using an
object-cloud.
[0048] FIG. 1 is an overall configuration diagram of a deep
learning-based object recognition system using a PIR sensor
according to the present invention. The PIR sensor detects an
infrared signal of an external object, and a deep learning-based
object recognition cloud recognizes an object and provides a result
of the recognition by a learning method using an artificial CNN and
through classification after learning.
[0049] An intrusion prevention system using a digital PIR sensor is
likely to erroneously sense an object rather than a human. To solve
this error, according to the present invention, an analog PIR
sensor and an object detection system using machine learning are
built.
[0050] As illustrated in FIG. 2A, an intrusion prevention system
using a PIR sensor of the related art senses a motion on the basis
of a threshold of the PIR sensor. When a digital logic value
exceeds a fixed threshold, it is determined that an intrusion
occurred and thus a logic HIGH value is generated, and otherwise,
it is determined that no intrusion has occurred and thus a logic
LOW value is generated.
[0051] In contrast, as illustrated in FIG. 2B, an analog PIR sensor
provides an output on the basis of various voltage scales within a
specific area without generating a binary output using a
threshold.
[0052] In the present invention, frequencies are generated by
performing a Fast Fourier Transform (FFT) on an analog signal
sample obtained using an analog PIR sensor and are used as feature
vectors of a CNN.
[0053] In the present invention, signal patterns of motions of a
human and an animal are learned using an artificial CNN to identify
whether a recognized object is a human or an animal.
[0054] As illustrated in FIG. 2B, the type of an object and a
distance to the object may be determined on the basis of the
difference in waveform between infrared rays from a human and an
animal.
[0055] A configuration of a deep learning-based object recognition
system using a PIR sensor according to the present invention will
be described in detail below.
[0056] FIG. 3 is a detailed configuration diagram of a deep
learning-based object recognition system using a PIR sensor
according to the present invention.
[0057] The deep learning-based object recognition system using a
PIR sensor according to the present invention includes a data
collection process executor 100 that performs a data collection
process of collecting values of the PIR sensor according to a
sampling period, a data classification process executor 200 that
performs a data classification process of inputting collected
values of the PIR sensor to a model of an artificial neural network
and transmits a result of the inputting the values to a cloud using
a RESTful API, and an object recognition cloud system 300 that
stores information received from the data classification process
executor 200 in a database, transmits the stored information when a
web application requests the information, and represents
information collected and classified by devices using the RESTful
API.
[0058] Here, the object recognition cloud system 300 includes a
RESTful AP framework for storing information received from the data
classification process executor 200 in the database and
transmitting the stored information when a web application requests
the information, and a monitoring web application for representing
information collected and classified by devices using the RESTful
API.
[0059] FIG. 4 is a flowchart of an event processing process of a
deep learning-based object recognition system using a PIR sensor
according to the present invention.
[0060] The data collection process executor 100 collects values of
the PIR sensor according to a sampling period and inputs the
collected values to a model of a trained artificial neural
network.
[0061] The trained artificial neural network classifies the input
data as one of three results: no object, an animal, and a human,
and represents a distance to a corresponding object.
[0062] A result of the classification and representation is
transmitted to the object recognition cloud system 300 through the
RESTful API.
[0063] The object recognition cloud system 300 forms a RESTful API
framework and a monitoring web application, and the RESTful API
framework stores information received from the data classification
process executor 200 in the database and transmits the information
when a web application requests the information.
[0064] The monitoring web application represents information
collected and classified by various devices through a RESTful API
provided by the RESTful API framework.
[0065] A deep learning-based object recognition method using a PIR
sensor according to the present invention will be described in
detail below.
[0066] FIG. 5 is a flowchart of a deep learning-based object
recognition method using a PIR sensor according to the present
invention.
[0067] In the deep learning-based object recognition method using a
PIR sensor according to the present invention, first, the object
data collection process executor 100 periodically receives the
amount of infrared rays measured by the PIR sensor (sampling rate:
100 Hz) (S501).
[0068] The object data collection process executor 100 collects
values of the PIR sensor at a rate of 100 Hz. When the number of
pieces of data collected reaches 300, a classification process is
created.
[0069] When the classification process is created, some (e.g.,
fifteen pieces) of the collected data is removed and data is
continuously collected. Similarly, when the number of pieces of
data collected reaches 300, a classification process is
created.
[0070] Next, the object data classification process executor 200
inputs collected values of the PIR sensor to an artificial neural
network and classifies the input values (S502) and transmits a
result of the classification to the object recognition cloud system
300 (S503).
[0071] The object recognition cloud system 300 stores and
represents data received from an object (S504), determines whether
an unauthorized object is present (S505), and reports the presence
or absence of an unauthorized object (S506).
[0072] FIG. 6 is a diagram illustrating a real-time data processing
process.
[0073] The data classification process executor 200 calculates an
artificial neural network and thus may overload a microprocessor.
Therefore, in the present invention, it is necessary to limit two
or more duplicate values so as not to be generated.
[0074] A calculation time t.sub.n of a data classification process
may be defined by:
t.sub.n=t.sub.after.sub.n-t.sub.prev.sub.n [Equation 1]
[0075] A limitation of the number of classification processes
according to time may be defined by:
t.sub.after.sub.n<t.sub.prev.sub.n+1 [Equation 2]
[0076] If a range of t.sub.n is approximately i.+-..alpha. ms, the
data collection process may be prevented from being duplicated when
a period of the data collection process is longer than i.+-..alpha.
ms, which is a maximum value of t.sub.n.
[0077] Therefore, in the present invention, a data collection
process is created at time intervals of i.+-..alpha. ms.
[0078] FIGS. 7A and 7B are diagrams for describing a deep
learning-based object recognition algorithm using a PIR sensor
according to the present invention.
[0079] The deep learning-based object recognition algorithm uses an
artificial neural network. Input data refers to a total of 300
pieces of sample data.
[0080] The artificial neural network preferably includes a neural
network for extracting data features using a CNN and a neural
network for identifying a correlation between features using a
recurrent neural network (RNN).
[0081] The neural network for extracting data features is trained
with several waveforms of infrared rays from a human and an animal,
which may be received by the PIR sensor. This neural network is
capable of coping with a change of data values due to noise and
environmental changes.
[0082] The neural network for identifying a correlation between
features identifies a correlation on the basis of several features
extracted from one data sample and derives a final result.
[0083] The final result represents a total of three pieces of
classification data: a human, an animal and no object; and a
distance m.
[0084] The distance m of the final result is continuous data and is
obtained according to a linear regression. The rest (human, animal,
and no object) of the final result should represent a result
corresponding to 0 or 1 according to the presence or absence of an
object and are obtained according to a logistic regression.
[0085] An object recognition algorithm according to the present
invention is configured as a combination of the linear regression
and the logistic regression.
[0086] Cost functions according to the linear regression and the
logistic regression are different, and a final cost function is
calculated by adding values of cost functions of the linear
regression and the logical regression.
Cost linear = ( y linear - y ^ linear ) 2 Cost logistic = - y
logistic log ( y ^ logistic ) - ( 1 - y logistic ) log ( 1 - y ^
logistic ) [ Equation 3 ] J ( .theta. ) = 1 m i = 1 m [ Cost linear
+ Cost logistic ] [ Equation 4 ] ##EQU00002##
[0087] In the deep learning-based object recognition system and
method using a PIR sensor according to the present invention
described above, whether a recognized object is a human or an
animal is identified by learning signal patterns of motions of a
human and an animal using an artificial neural network (CNN) so
that the type of the object and a distance to the object may be
identified on the basis of the difference in waveform between
infrared rays from a human and an animal.
[0088] A deep learning-based object recognition system and method
using a PIR sensor according to the present invention has the
following advantages.
[0089] First, a human and an object can be identified using a new
type of PIR sensor and machine learning-based object detection
algorithm.
[0090] Second, object recognition performance can be improved by
extracting various signals obtained by the PIR sensor and
processing PIR data to extract a frequency component of a signal as
a feature vector.
[0091] Third, object recognition performance can be improved
through a learning method using an artificial convolutional neural
network (CNN) and classification after learning, and object
recognition performance and usability can be improved by building a
real-time-web-based monitoring system using an object-cloud.
[0092] Fourth, a false-alarm rate can be reduced using an object
recognition technology of measured data using deep learning.
[0093] It will be understood that the present invention may be
implemented in modified forms without departing from the essential
characteristics of the present invention as described above.
[0094] Therefore, the embodiments set forth herein should be
considered in a descriptive sense only and not for purposes of
limitation, the scope of the present invention is defined not by
the above description but by the appended claims, and all
differences within the scope of the present invention should be
construed as being included in the present invention.
* * * * *