![](bg7.png)
detections (true positives), even if at some individual
frames, their confidence is not so high or are missed
(false negatives). To help in this mission, we could
also incorporate tracking methods to analyze image
regions where previously there was a detection with
high accumulated confidence. It should be considered
that such solutions would come at the cost of increas-
ing the detection time, although it could be mitigated
if the sensor and the system can cope with a sufficient
frame rate.
The different performance of fire and smoke de-
tection is also worth studying further by examining
the composition of the datasets and maybe training
a separate method for each. Since smoke may ex-
pand faster than flames, it can provide an earlier clue
for the detection task, so improving its performance
is highly relevant. Also, newer real-time detectors
like YOLOR (Wang et al., 2021) are showing higher
speed than YOLOv4 while keeping accuracy, which
could favor its deployment to embedded devices in
industrial settings. This system was released after the
course of this investigation, so it is saved for future
work.
When deploying a fire detection system like the
one in this paper, one must consider the various
ethical questions related to camera-based detection,
due to humans potentially appearing in the footage.
Whenever a camera is capturing or streaming such
type of data to a remote location, privacy, security,
and GDPR concerns emerge. These concerns would
be significantly counteracted via edge computing,
with data processed as close as possible to where it
is being captured, diminishing transmission of sensi-
tive data to a different location through data networks.
In this regard, edge devices usually have fewer com-
puting capabilities, which is the reason why we are
aiming at deploying our system to suitable hardware,
such as NVIDIA Jetson nano. This also connects with
using detectors with low latency, such as YOLOR, as
mentioned above. Also, necessary frames must be
deleted as soon as computations are done. The present
system only uses one frame, but even combining sev-
eral frames with sufficient frame rate would mean that
the necessary data to be processed only affects a few
milliseconds of footage. Handling the data in this way
means that no sensitive data would ever be stored, or
transmitted elsewhere.
ACKNOWLEDGEMENTS
This work has been carried out by Otto Zell and
Joel P
˚
alsson in the context of their Bachelor Thesis
at Halmstad University (Computer Science and En-
gineering), with the support of HMS Networks AB
in Halmstad. Authors Hernandez-Diaz and Alonso-
Fernandez thank the Swedish Research Council (VR)
and the Swedish Innovation Agency (VINNOVA) for
funding their research.
REFERENCES
Abdusalomov, A., Baratov, N., Kutlimuratov, A., and
Whangbo, T. K. (2021). An improvement of the fire
detection and classification method using yolov3 for
surveillance systems. Sensors, 21(19).
Accosta, R. and Martin, D. (2017). Smoke detector spacing
for high ceiling spaces. Technical report, National Fire
Protection Association, NFPA.
Ahrens, M. (2021). Smoke alarms in us home fires. Tech-
nical report, National Fire Protection Association,
NFPA.
Avazov, K., Mukhiddinov, M., Makhmudov, F., and Cho,
Y. I. (2022). Fire detection method in smart city envi-
ronments using a deep-learning-based approach. Elec-
tronics, 11(1).
Bochkovskiy, A., Wang, C.-Y., and Liao, H.-Y. M. (2020).
Yolov4: Optimal speed and accuracy of object detec-
tion.
Bu, F. and Gharajeh, M. S. (2019). Intelligent and vision-
based fire detection systems: A survey. Image and
Vision Computing, 91:103803.
Fonollosa, J., Sol
´
orzano, A., and Marco, S. (2018). Chem-
ical sensor systems and associated algorithms for fire
detection: A review. Sensors, 18(2).
Gonzalez, A., Zuniga, M. D., Nikulin, C., Carvajal, G.,
Cardenas, D. G., Pedraza, M. A., Fernandez, C. A.,
Munoz, R. I., Castro, N. A., Rosales, B. F., Quinteros,
J. M., Rauh, F. J., and Akhloufi, M. A. (2017). Ac-
curate fire detection through fully convolutional net-
work. In 7th Latin American Conference on Net-
worked and Electronic Media (LACNEM 2017), pages
1–6.
HMS (2022). https://www.hms-networks.com.
Li, Z., Peng, C., Yu, G., Zhang, X., Deng, Y., and Sun,
J. (2017). Light-head r-cnn: In defense of two-stage
object detector.
Muhammad, K., Ahmad, J., Lv, Z., Bellavista, P., Yang, P.,
and Baik, S. W. (2019). Efficient deep cnn-based fire
detection and localization in video surveillance appli-
cations. IEEE Transactions on Systems, Man, and Cy-
bernetics: Systems, 49(7):1419–1434.
Nilsson, F., Jakobsen, J., and Alonso-Fernandez, F. (2020).
Detection and classification of industrial signal lights
for factory floors. In 2020 International Conference
on Intelligent Systems and Computer Vision (ISCV),
pages 1–6.
Tan, M., Pang, R., and Le, Q. V. (2020). Efficientdet: Scal-
able and efficient object detection. In 2020 IEEE/CVF
Conference on Computer Vision and Pattern Recogni-
tion (CVPR), pages 10778–10787, Los Alamitos, CA,
USA. IEEE Computer Society.
Image-Based Fire Detection in Industrial Environments with YOLOv4
385