Towards automatic smoke detector inspection: Recognition of the smoke detectors in industrial facilities and preparation for future drone integration

Lukas Kratochvila; Jakub Stefansky; Simon Bilik; Robert Rous; Tomas Zemcik; Michal Wolny; Frantisek Rusnak; Ondrej Cech

cs.CVMar 25, 2026

Towards automatic smoke detector inspection: Recognition of the smoke detectors in industrial facilities and preparation for future drone integration

Authors: Lukas Kratochvila, Jakub Stefansky, Simon Bilik, Robert Rous, Tomas Zemcik, Michal Wolny, Frantisek Rusnak, Ondrej Cech,

Organizations: Department of Control and Instrumentation, Brno University of Technology, Brno, Czech Republic · Department of Cybernetics and Biomedical Engineering, VSB - Technical University of Ostrava, Ostrava, Czech Republic · Institute for Research and Applications of Fuzzy Modeling, University of Ostrava, Ostrava, Czech Republic · Department of Informatics, Mendel University in Brno, Brno, Czech Republic ·

Abstract

Fire safety consists of a complex pipeline, and it is a very important topic of concern. One of its frontal parts are the smoke detectors, which are supposed to provide an alarm prior to a massive fire appears. As they are often difficult to reach due to high ceilings or problematic locations, an automatic inspection system would be very beneficial as it could allow faster revisions, prevent workers from dangerous work in heights, and make the whole process cheaper. In this study, we present the smoke detector recognition part of the automatic inspection system, which could easily be integrated to the drone system. As part of our research, we compare two popular convolutional-based object detectors YOLOv11 and SSD widely used on embedded devices together with the state-of-the-art transformer-based RT-DETRv2 with the backbones of different sizes. Due to a complicated way of collecting a sufficient amount of data for training in the real-world environment, we also compare several training strategies using the real and semi-synthetic data together with various augmentation methods. To achieve a robust testing, all models were evaluated on two test datasets with an expected and difficult appearance of the smoke detectors including motion blur, small resolution, or not complete objects. The best performing detector is the YOLOv11n, which reaches the average [email protected] score of 0.884. Our code, pretrained models and dataset are publicly available.

Figures & tables

Dataset	Real-Normal	Real-Difficult	Gen-Real	Gen-Render
Train ( R + G )	1004	-	1536	1536
Valid ( R + G )	334	-	384	384
Test ( R + R )	334	127	-	-

TABLE I : Image counts and portions of real and generated data in the experimental datasets.

Hyperparameter	YOLOv11	RT-DETRv2	SSD
Image size	640x640	640x640	300x300
Number of epochs	120	120	120
Batch size	16	16	16
Optimizer	AdamW	AdamW	SGD
Learning rate	$7\times 10^{-4}$	$2\times 10^{-5}$	$1.5\times 10^{-3}$
Weight decay	$5\times 10^{-4}$	$2\times 10^{-4}$	$2\times 10^{-5}$

TABLE II : Training hyperparameters of the selected models.

rPi type	CPU	Core	Frequency [MHz]	GPU
Raspberry Pi 4B	Broadcom BCM2711	Cortex-A72	1500	VideoCore VI
Raspberry Pi 5	Broadcom BCM2712	Cortex-A76	2400	VideoCore VII

TABLE III : Parameters of the evaluated rPi embedded computers

Training dataset	Validation dataset	[email protected] Validation	[email protected] Test- N	[email protected] Test- D	[email protected] Test Avg	[email protected] Test Diff
YOLOv11m
Real	Real	0.941	0.934	0.688	0.811	0.246
Real	Generated	0.932	0.864	0.331	0.597	0.533
Generated	Generated	0.995	0.495	0.276	0.385	0.219
Generated	Real	0.658	0.578	0.372	0.475	0.206
Mixed	Real	0.918	0.924	0.731	0.827	0.193

TABLE IV : Detection results over various datasets.

YOLOv11 over the mixed training and real validation dataset.
Model variant	[email protected] Validation	[email protected] Test- N	[email protected] Test- D	[email protected] Test Avg	[email protected] Test Diff
YOLOv11n	0.942	0.940	0.828	0.884	0.112
YOLOv11s	0.945	0.934	0.776	0.855	0.158
YOLOv11m	0.918	0.924	0.744	0.834	0.180
YOLOv11x	0.943	0.933	0.803	0.868	0.130
SSD over the real training and validation datasets.

TABLE V : Detection results over various variants of object detectors.

Augmentations	[email protected] Test- N	[email protected] Test- D	Test Avg
YOLOv11 over the mixed training and real validation dataset.
Default	0.940	0.828	0.884
Albumentations	0.936	0.815	0.875
SSD over real training and validation dataset.
Default	0.909	0.843	0.876
Albumentations	0.916	0.671	0.793

TABLE VI : Effect of augmentations on evaluated object detectors.

Fig. 2 : Structure of the proposed ROS-2 pipeline.

rPi type	preprocess [ms]	interference [ms]	postprocess [ms]	total inference time [ms]	possible FPS rate
YOLOv11n
Raspberry Pi 4B	9.8	597.5	2.1	609.4	1.640
Raspberry Pi 5	3.1	153.5	0.5	157.1	6.365
SSD with VGG16 backbone
Raspberry Pi 4B	40.5	4414.9	0.8	4456.2	0.224
Raspberry Pi 5	3.7	1175.0	0.3	1179.0	0.848

TABLE VII : Inference times comparison of YOLOv11n object detector on rPi systems using the image of 320x320 px.

Fig. 3 : Low confidence YOLOv11n detections.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Attention-Based Spatio-Temporal Approach for Fire and Smoke Detection
Oct 14, 2026Alireza Mirrashid, Mohammad Khoshbin, Ali Atghaei +1
VEUNEX, Germany
Impact of Dataset Composition on Embedded Real-Time UAV Wildfire Detection Using Compact YOLO Models
Aug 1, 2026Eduardo de los Santos, Andre S. Kelbouscas, Ricardo B. Grando +1Synthetic Data Augmentation Wildfire Detection
Robotics and AI Lab, Technological University of Uruguay, Rivera, Uruguay
Intelligent Spatial Estimation for Fire Hazards in Engineering Sites: An Enhanced YOLOv8-Powered Proximity Analysis Framework
Mar 10, 2026Ammar K. AlMhdawi, Nonso Nnamoko, Alaa Mashan Ubaid
University of Greater Manchester, A676 Deane Rd, Bolton, BL3 5AB, , UK · Edge Hill University, St Helens Rd, Ormskirk, L39 4QP, , UK · University of Khorfakkan, Al Jaradia , Sharjah, 887W+2P2, , UAE

Oct 14, 2026J/K move · Enter open · S save

Attention-Based Spatio-Temporal Approach for Fire and Smoke Detection

Alireza Mirrashid, Mohammad Khoshbin, Ali Atghaei, Hassan Shahbazi

VEUNEX, Germany

Open paper AskPDF arXiv

In various industries, smoke and fire are two of the most important threats in the workplace. One of the common methods for detecting smoke and fire is the use of infrared thermal and smoke sensors, which cannot be used in outdoor applications. Therefore, the use of vision-based methods seems necessary. The problem of smoke and fire detection is spatiotemporal and requires spatiotemporal solutions. This paper presents a method that uses spatial features along with temporal-based features to detect smoke and fire in the scene. It consists of three main parts; the task of each part is to reduce the error of the previous part so that the final model has a robust performance. This method also uses transformer modules to increase the accuracy of the model. The results of our model show the proper performance of the proposed approach in solving the problem of smoke and fire detection and can be used to increase workplace safety.

Abstract

Object Detection

Explore similar work

Attention-Based Spatio-Temporal Approach for Fire and Smoke Detection

Impact of Dataset Composition on Embedded Real-Time UAV Wildfire Detection Using Compact YOLO Models

Intelligent Spatial Estimation for Fire Hazards in Engineering Sites: An Enhanced YOLOv8-Powered Proximity Analysis Framework