Department and other emergencies all require thoseDepartment and other emergencies all require those

Department of Information Technology, Universitas
Sumatera Utara, Medan, Indonesia


Emergency alert systems serve as a critical link in the
chain of crisis communication, and they are essential to minimize loss during
emergencies. Acts of terrorism and violence, chemical spills, amber alerts,
nuclear facility problems, weather-related emergencies, flu pandemics, and
other emergencies all require those responsible such as government officials,
building managers, and university administrators to be able to quickly and
reliably distribute emergency information to the public. This paper presents
our design of a deep-learning-based emergency warning system. The proposed system
is considered suitable for application in existing infrastructure such as
closed-circuit television and other monitoring devices. The experimental
results show that in most cases, our system immediately detects emergencies
such as car accidents and natural disasters.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!

order now

Emergency alert system; Deep-learning;

Abstrak. Sistem peringatan darurat
berfungsi sebagai jalur kritis dalam rantai komunikasi krisis, dan penting
untuk meminimalkan kerugian selama keadaan darurat. Tindakan terorisme dan
kekerasan, tumpahan kimia, peringatan kuning, masalah fasilitas nuklir, keadaan
darurat terkait cuaca, pandemi flu, dan informasi darurat darurat lainnya
kepada publik. Jurnal ini menyajikan desain sistem peringatan darurat berbasis
pembelajaran yang mendalam. Sistem yang diusulkan dianggap cocok untuk aplikasi
di infrastruktur yang ada seperti televisi sirkuit tertutup dan perangkat
pemantauan lainnya. Hasil percobaan menunjukkan bahwa dalam banyak kasus,
sistem kami segera mendeteksi keadaan darurat seperti kecelakaan mobil dan
bencana alam.

Kata Kunci:
Bencana, Deep-learning, Sistem Siaga Darurat.


The Emergency Alert System (EAS) is a national
warning system that was implemented in the United States on November 29, 1997.
However, recent years have seen the exposure of a number of drawbacks of the
EAS. For example, systems that rely on the use of cellular phone 1,2 or radio
broadcast 3 networks are often unable to reach individuals who are located
inside of buildings. The interiors of many buildings at universities, research
centers, office complexes, manufacturing plants, and other locations often have
very poor radio and cellular phone reception because of interference caused by
equipment located within the building, or because of a shielding effect created
by the building structure itself.

In addition, current emergency systems are not easily
able to reach the right people, in the right location, at the right time.
Although services relying on Cellular phones services, text messaging services,
and e-mail services can target specific individuals, but they will would not be
effective for a location-specific emergency because such these services are
only able to target individuals people selectively on an individual basis by
phone number or e-mail address, regardless of their physical location. Sirens
4 can provide a quick alert, but they may not yield desired results because
the sound may not reach all locations, and some individuals in some areas may
ignore a siren that provides no specific information about the emergency.
People relying solely on cellular telephones would be excluded from the
warning. Networks like such as Ethernet and WiFi are prone to failure in times
of an emergency due to because potential power outages could shutting down the
network or one or more network devices, thereby causing the communication
failure with an entire building or geographical area to fail.

Closed-circuit television (CCTV), also known as an
emergency or crime monitoring system, involves the use of video cameras to transmit
a signal to a specific place, on a limited set of monitors. It differs from
broadcast television in that the signal is not publicly transmitted, although
it may employ point-to-point (P2P), point-to-multipoint, or mesh wire-less
links. Lately, CCTV technology has been enhanced with a shift toward
Internet-based products and systems 5,6 and other technological developments.
For instance, a CCTV at an ATM could be used to capture a user’s PIN as it is
entered via the keypad, without their knowledge. The devices are small enough
not to be noticed, and are positioned such that they enable others to monitor
the keypad. Images may be transmitted wirelessly to the criminal.

The field of artificial intelligence (AI) is in a
period of unprecedented improvement. Deep-learning technology 7–9 is behind
most of the recent breakthroughs in object recognition, natural language
processing, and speech recognition. This paper introduces an emergency alert
system based on the use of deep-learning technology. The main advantage of the
pro-posed system is that it does not require additional devices or infrastructure.
We adapt a deep-learning-based real-time video analyzing module inside the CCTV
device. Our system uses computer simulation to immediately detect accidents and
natural disasters. The remaining part of this paper is organized as follows. We
detail our proposed emergency alert system in Section 2. Experiments and
measurement results are provided in Section 3. Finally, we conclude the paper
in Section 4.

Emergency Alert System

Convolutional neural
networks 10,11 are hierarchical machine-learning models capable of learning a
complex representation of images using vast amounts of data. They are in-spired
by the human visual system and learn multiple layers of transformations, which
extract a progressively more sophisticated representation of the input.

Our system uses heuristic/knowledge-based
machine-learning technology, in which the process to generate descriptors
starts with an early step to discover all the labels included in our problem
domain (knowledge pool). This is done by processing all the available video images
with the analyzer API, and then removing the duplicated label values. The
proposed analyzer uses basic Deep Neural Network (DNN) 12 architectures for
object detection and parsing to generate compositional models. This step
resembles a codebook or dictionary generation process. Each input image is then
encoded as a list of probability values whose length is determined by the size
of the dictionary. The ith value in this descriptor corresponds to the
probability returned by the analyzer for the ith label, or zero otherwise. Fig.
1 shows the various steps of the real-time video analyzing process. Our system
has the following benefits:

•           Scalable: DNNs 13,12 scale to billions
of parameters giving them the capacity to learn highly complex concepts and
thousands of categories. Modern hardware and an abundance of data enable our
system to train larger and more powerful networks.

•           Fast: the trained model stores its
knowledge compactly in learned parameters, making it easy to deploy in any
environment. There is no need to store any additional data to make predictions
for new inputs. This means that we could easily use them in embedded devices
such as CCTV to provide responses in milliseconds.

•           Flexible: unlike traditional computer
vision approaches 14 our system learns to extract discriminative features
from the input using the provided training data, instead of using hand-engineered
feature extractors such as SIFT 15 and LBP 16. This makes these features
easy to adapt to problems in any domain.

We evaluated our system
by developing a simulator and use two different types of disasters for
experimentation purposes. The first is a fire and the other is a car accident.
We per-formed the simulation by using the following system: Intel Core
i7-4810MQ (4 Cores -8 Threads), Nvidia Quadro K4100M (4 GB) GPU, and 16 GB RAM.
The use of a GPU to train deep neural networks enabled us to increase the
runtime, ranging from 3x faster to 15x faster 17,18. Our simulator is programmed
with Python. The simulator randomly generates an emergency situation based on the
Poisson process 19. We simulated the two emergency situations more than 100
times at each measuring point while progressively increasing the level of
strength (weak, normal, strong) and the lambda value (? = 0.1, 0.3, 0.6).
Values of ? = 0.1 and 0.6 mean an event (dis-aster) occurred on average every
10 and 5 s, respectively. In Figs. 2 and 3, lines indicated with various colors
have different meanings. Each line was plotted using data from our database.







Fig. 2 shows that the analyzer normally detects a car accident within
600 ms. In this figure, the dark-blue line means a car invaded the yellow line
and collided with a car traveling in the opposite direction. Disasters that
were generated by the system disappeared after 1.5 s. Fig. 3 shows the
monitoring results for a fire situation with each curve showing the level of
the unexpected situation. In this case, three or four factors (lines) indicate
the fire situation.

Tables 1 and 2 present the experimental parameters and results for a
fire and car accident, respectively. In Table 2, the term avg. detection time
refers to the time that passed after the event occurred, whereas accuracy refers
to the probability of detecting events. The results show that the proposed
system can detect fire situations around 99% of the time within 400 ms, which
means that our system can easily detect fire situations. How-ever, in the case
of a car accident, the detection accuracy of our system is below 96% because it
is quite difficult and requires time to recognize an “accident” or “collision”.
Therefore, our system has the ability to detect fire outbreaks more rapidly
than situations involving car accidents. This indicates that our sys-tem would
need to accumulate a certain amount of knowledge by using a heuristic process
to enable it to detect car accidents more accurately.


The emergency alert system we proposed is designed to be suitable for
natural disaster detection. The proposed system uses deep-learning technology
to detect and analyze disasters. We carried out experimental measurements to
assess the performance of our proposed system while increasing the disaster
strength and event frequency. The evaluation showed that the average detection
time and accuracy for situations involving “fire” demonstrated a higher
detection rate than for those involving “car accidents”. These experimental
results could be applied in practice by adapting our EAS system to real CCTV or
other monitoring devices. Although our computer simulation only generated the
above-mentioned two types of emergency events, in real environments, there
exist many different emergency situations, not only disasters but also various
types of criminal situations.


1   V.H. Soragavi, S. Balundagi, S. Mulla, S. Kumbalavati, Droid alert
system using 3g network.

2 L. Hariprasath, R. Dhivya, S. Adithya, C. Kazhipattur, Emergency alert
system using android, IJREAT Int. J. Res. Eng. Adv. Technol. 1 (1).

3   M.-S. Baek, Y.-H. Lee, G. Kim, S.-R.
Park, Y.-T. Lee, Development of T-DMB emergency broadcasting system and
trial service with the legacy receivers, IEEE Trans. Consum.
Electron. 59 (1) (2013) 38–44.

4 K. Richard Ridderinkhof, B.U. Forstmann,
S.A. Wylie, B. Burle, W.P. van den Wildenberg, Neurocognitive mechanisms
of action control: resisting the call of the sirens, Wiley
Interdiscip. Rev. Cogn. Sci. 2 (2) (2011) 174–192.

5 H. Kruegle, CCTV Surveillance: Video
Practices and Technology, Butterworth-Heinemann, 2011.

6 D. Choi, N. Baik, CCTV monitoring technology of nighttime low-light
situations at the ITS Center in Korea, in: 20th ITS World Congress, 2013.


7   L. Deng, G. Hinton, B. Kingsbury, New
types of deep neural network learning for speech recognition and related
applications: An overview, in: 2013 IEEE International Conference on
Acoustics, Speech and Signal Processing, ICASSP, IEEE, 2013, pp.

8   J. Ngiam, A. Coates, A. Lahiri, B.
Prochnow, Q.V. Le, A.Y. Ng, On optimization methods for deep learning, in:
Proceedings of the 28th International Conference on Machine Learning (ICML-11),
2011, pp. 265–272.

9   I. Sutskever, J. Martens, G. Dahl, G.
Hinton, On the importance of initialization and momentum in deep learning, in:
Proceedings of the 30th International Conference on Machine Learning, ICML-13,
2013, pp. 1139–1147.

A. Krizhevsky,
I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional
neural networks, in: Advances in Neural Information Processing Systems, 2012,
pp. 1097–1105.

S. Ji, W. Xu, M. Yang, K. Yu, 3d
convolutional neural networks for human action recognition, IEEE Trans. Pattern
Anal. Mach. Intell. 35 (1) (2013) 221–231.

G. Hinton, L. Deng, D. Yu, G.E. Dahl,
A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T.N.
Sainath, et al., Deep neural networks for acoustic modeling in
speech recognition: The shared views of four research groups, IEEE Signal
Process. Mag. 29 (6) (2012) 82–97.

K. Vesely,` A.
Ghoshal, L. Burget, D. Povey, Sequence-discriminative training of deep neural
networks., in: INTERSPEECH, 2013, pp. 2345–2349.

Y. Shirai, Three-dimensional Computer
Vision, Springer Science & Business Media, 2012.

M. Niemeijer,
M.K. Garvin, K. Lee, B. van Ginneken, M.D. Abramoff,` M. Sonka, Registration of
3D spectral OCT volumes using 3D SIFT feature point matching, in: SPIE Medical Imaging, International Society for Optics
and Photonics, 2009, pp. 72591I–72591I.

J. Shelton, G. Dozier, K. Bryant, J.
Adams, K. Popplewell, T. Abegaz, K. Purrington, D.L. Woodard, K. Ricanek,
Genetic based LBP feature extraction and selection for facial
recognition, in: Proceedings of the 49th Annual Southeast Regional Conference, ACM,
2011, pp. 197–200.


A. Kerr, G. Diamos, S. Yalamanchili,
Modeling GPU-CPU workloads and systems, in: Proceedings of the 3rd
Workshop on General-Purpose Computation on Graphics Processing
Units, ACM, 2010, pp. 31–42.

M. Arora, S. Nath, S. Mazumdar, S.B.
Baden, D.M. Tullsen, Redefining the role of the CPU in the era of CPU–GPU
integration, IEEE Micro. 32

(6) (2012) 4–16.

W. Fischer, K. Meier-Hellstern, The
Markov-modulated Poisson process (MMPP) cookbook, Perform. Eval. 18
(2) (1993) 149–171.