You are on page 1of 2

2019 IEEE International Conference on Consumer Electronics

- Taiwan (ICCE-TW)


New Generation Artificial Intelligent Vending Machine System
based on LoRaWan IOT Network
Chih-Chun Hsu1, Yi-Chang Lin2, Yeong-Long Shiue1, and Chi-Chia Sun1, Member, IEEE
1
Department of Electrical Engineering, National Formosa University, Yunlin County, Taiwan
2
Far Easy Pass LTD., New Taipei City, Taiwan

to identify the structure of the buyer and the LoRaWan IOT.


ABSTRACT On the other hand, the distribution of the source of individual
The Internet of Things and artificial intelligence are one of vending machines can also be stored for analysis.
the mainstream of technological development in this era.
Machine Learning now is often used for image recognition, II. SYSTEM ARCHITECTURE
and it is necessary to develop different models according to In this paper, in order to monitor individual vending
requirements in the past. This paper combines Artificial
machine in public or many of them, wide-area Internet of
Intelligence and Internet of Things into the vending system,
Things technology is applied. The core system is based on
and system is presented in the form of ticket vending
LoRaWAN architecture [4], consists of sensing terminal, a
machines and veterans. The proposed system can identify
the customer's age, gender data, and then sends the data to gateway, a server and a monitoring terminal as shown Fig. 1.
cloud server through the LoRaWan IOT Network to The vending machine is equipped within an embedded system
analyze the customers and merchandise sales. It will with a LoRa module and a camera The camera will capture the
monitor the current status of the vending machine at the image of the buyer, then it will perform gender identification
same time. and age recognition within AI technology. When the gateway
receives the data from the vending machine, it will transmit to
Keywords: AI, Vending Machine, LoRaWan, CNN, IOT the remote back-end server via TCP/IP for data processing and
statistics. A remote monitoring system corresponds to the
I. INTRODUCTION application layer at LoRaWAN. In this part, a monitoring data
Vending machines have changed the traditional shopping interface is implemented, and the data is obtained from the
model. The traditional shopping mode requires storefronts and server, so that the user can view the vending machine through
personnel, but the vending machines are not restricted to the the monitoring interface.
venue. The existence of vending machines can be found on the
street, on campus, and in the building. Manpower support is
required for the goods, which saves the resources spent on
opening the store, such as physical store rent, personnel costs,
storage space, etc. [1]. Due to the limitations of the vending
machine volume, unlike convenience stores, which can sell all
kinds of items, most of them are targeted to sell snacks,
microwave meals or drinks. Therefore, in order to maximize
the commercial value of a single vending machine, we have to
maximize the commercial value [2, 3]. Find out which items
are popular and adjust for the merchandise. When purchasing
goods, customers will choose according to their own Fig. 1. System Schematic
preferences. However, if you zoom in on the ethnic groups,
you will find that there are similar preferences in the ethnic The buyer identification system in the vending machine
groups divided by certain conditions. For example, in terms of adopted a deep learning model based on CNN networks for
drinks, using age as a condition, you can find most children age identification and gender identification. After learning
don't like bitter coffee, sugar-free tea, they prefer soft drinks, many low-order features, it is learned by high-order features
juices and other sugar-sweetened drinks. Young-aged or composed of low-order features [5]. The deep learning
middle-aged people have a higher chance of choosing coffee network is used for age-level discrimination and gender
and sugar-free. In addition to age, gender may be one of the discrimination required training itself. In system, once there is
reasons for choosing different commodities, even some a new transaction been triggered, vending machine will take a
products are produced for specific gender. picture, recognizes the face block, resizing after the face block
This paper presented a value-added structure for vending is cut off for identification. After the image size is adjusted,
machines, combined with the CNN facial recognition system two training filters in the model are used to perform two
convolutions, then thresholding and block histogram statistics,

978-1-7281-3279-2/19/$31.00
Authorized licensed use limited to: University of Exeter. Downloaded on May 06,2020 at 10:35:17 ©2019
UTC from IEEE Xplore. Restrictions apply.IEEE
and finally placed in the trained models for classification. Table I
CIA Database Content
III. EXPERIMENTAL RESULTS Gender Age
Category male female <12 13~19 20~45 >45
The hardware of ticket vending machine includes embedded CIA 1,080 1,008 313 592 934 248
computer, RFID card reader, ticket printer, touch screen, Morph 46,645 8,489 16,632 14,141 15,609 8,752
camera module and LoRa module. Due to the size and
performance of the transaction, the embedded computer is Three CNN models were used for training, namely AlexNet,
selected as a control core of the vending machine. Wiring each VGG-16, and GoogLeNet. Three training methods were also
device with the computer and wiring it to complete a vending used, namely SGD, Adagrad, and Adam. After classification,
machine. The proposed vending machine is shown in Fig. 2. the test results were tested. Finally, the best collocation
method is applied to the machine, as in Table II.
Table II
Testbench Results
Database Method Age Gender
AlexNet(SGD) +
VGG-16(Adam) +
CIA 76.02% 90.61%
GoogLeNet(Adam) with
Voting
Fig. 2. Ticket Machine Entity and its user interface. AlexNet(SGD) +
VGG-16(Adam) +
Morph 76.68% 98.52%
The human-machine interface provides a touch panel for the GoogLeNet(Adam) with
Voting
user to control and purchase the product. The execution flow is
shown as Fig. 3. After the program is started, the settings
required by the system will be loaded. Once RFID card reader
and printer are detected on-line, it will check the connection. If CONCLUSION
it is not detected, it will inform the error, then the interface In this paper, we combined LoRaWan and AI technology to
program will be opened and wait for the button for click. It can form a sales system for smart vending machine. It can
be divided into three functions: ticket data inquiry, shopping, understand vending machine’s buyers and markets. Face
and item modification. recognition is applied to analyze the main source of the
vending machine in order to adjust the sales, thereby achieving
Open program a higher profit.

Load settings
REFERENCES
[1] A. Ramzan, S. Rehman, and A. Perwaiz, "RFID
technology: Beyond cash-based methods in vending
Detecting printers machine," International Conference on Control and
and card readers
Robotics Engineering, 2017, pp. 189-193.
Alert printer or card
reader connection
[2] G. Nemoto and K. Hiraishi, "Modeling and optimization of
failed item changes in vending machines,"Asian Control
Open the GUI
then wait Conference, 2017, pp. 19-24.
[3] R. Gruen and E. Liang, "NuiVend - Next Generation
Vending Machine," International Conference on
Follow the button to execute Computational Science and Computational Intelligence,
different projects
2016, pp. 545-548.
[4] LoRa Alliance, "LPWA Technologies-Unlock New IoT
Market Potential," ed, 2015.
Electronic ticket
Ticketing process
Item modification
End program
[5] A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson,
inquiry process process
"CNN Features Off-the-Shelf: An Astounding Baseline for
Recognition," IEEE Conference on Computer Vision and
Fig. 3. Ticket Machines HMI System Flowchart. Pattern Recognition Workshops, 2014, pp. 512-519.
[6] K. Ricanek and T. Tesafaye, "MORPH: a longitudinal
The training databases include CIA and Morph, as in Table image database of normal adult age-progression,"
I. CIA is a face database of Taiwanese and Chinese. The age International Conference on Automatic Face and Gesture
distribution ranges from 6 to 80 years old. The total samples Recognition, 2006, pp. 341-345.
are 2,088; Morph is a Western-based face database with age
distribution between 16 and 77. The total samples are 55,134
[6].

Authorized licensed use limited to: University of Exeter. Downloaded on May 06,2020 at 10:35:17 UTC from IEEE Xplore. Restrictions apply.

You might also like