Next Article in Journal
Effectiveness of Surface Pre-Application of Compressed Air Foam in Delaying Combustion Spread to Adjacent Buildings
Previous Article in Journal
Comparative Analysis of Chemical Reaction Mechanisms of Ammonia-n-Heptane Mixtures: From Ignition, Oxidation, and Laminar Flame Propagation to Engine Applications
Previous Article in Special Issue
Regression Analysis of Heat Release Rate for Box-Type Power Bank Based on Experimental and Machine Learning Methods
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Fire
Volume 8
Issue 9
10.3390/fire8090358
fire-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Review

Large-Space Fire Detection Technology: A Review of Conventional Detector Limitations and Image-Based Target Detection Techniques

by
Li Deng
1,2,3,
Siqi Wu
1,
Shuang Zou
4 and
Quanyi Liu
1,2,3,*
1
College of Civil Aviation Safety Engineering, Civil Aviation Flight University of China, Guanghan 618307, China
2
Sichuan Key Laboratory of Civil Aircraft Fire Science and Safety Engineering, Civil Aviation Flight University of China, Guanghan 618307, China
3
Sichuan All-Electric Aviation Aircraft Key Technology Engineering Research Center, Guanghan 618307, China
4
Airport Operation Security Department, Suining Branch, Civil Aviation Flight University of China, Suining 629000, China
*
Author to whom correspondence should be addressed.
Fire 2025, 8(9), 358; https://doi.org/10.3390/fire8090358
Submission received: 27 July 2025 / Revised: 31 August 2025 / Accepted: 4 September 2025 / Published: 7 September 2025
(This article belongs to the Special Issue Building Fire Dynamics and Fire Evacuation, 2nd Edition)
Browse Figures
Versions Notes

Abstract

With the rapid development of large-space buildings, their fire risk has become increasingly prominent. Conventional fire detection technologies are often limited by spatial height and environmental interference, leading to false alarms, missed detections, and delayed responses. This paper reviews 83 publications to analyze the limitations of conventional methods in large spaces and highlights the advantages of and current developments in image-based fire detection technology. It outlines key aspects such as equipment selection, dataset construction, and target recognition algorithm optimization, along with improvement directions including scenario-adaptive datasets, model enhancement, and adaptability refinement. Research demonstrates that image-based technology offers broad coverage, rapid response, and strong anti-interference capability, effectively compensating for the shortcomings of conventional methods and providing a new solution for early fire warning in large spaces. Finally, future prospects are discussed, focusing on environmental adaptability, algorithm efficiency and reliability, and system integration, offering valuable references for related research and applications.

1. Introduction

In China, the conceptualization of large-space buildings is formally defined within authoritative references, including the Encyclopedia of Chinese Civil Engineering and Architecture and the industry standard Technical specification for large-space intelligent active control sprinkler systems (CECS 263:2009) [1]; specifically, Clause 2.1.1 of CECS 263:2009 explicitly classifies civil and industrial buildings with an internal clear height exceeding 8 m and warehouse buildings with an internal clear height exceeding 9 m as large-space buildings [2], while the Encyclopedia of Chinese Civil Engineering and Architecture, alongside established architectural principles, further categorizes large-span public buildings—such as sports facilities, performance venues, exhibition centers, and transportation terminals—within this classification due to their functional requirement for continuous, column-free internal spaces.
Since the beginning of the 21st century, the construction industry has developed at an extremely rapid pace, with a continuous emergence of various types of large-space buildings, such as high-rise residential buildings, large-scale public entertainment venues, large warehouses, hangars, waiting halls, and large-scale underground parking lots. According to the composition of the completed building area by national construction enterprises in 2021, the proportion of residential buildings is the largest. The data indicate that the proportion of completed residential building area was 66.26% in 2021, followed by the proportion of completed factory and building area, which accounted for 13.81%. With the increasing number of large-space buildings, the probability of fires occurring in such structures is also on the rise [3]. In recent years, the number of fires, the number of fatalities, and direct property losses have generally increased. The devastation caused by fires is alarming and has inflicted incalculable losses on our nation and society. With the rapid development of our economy and the high concentration of social material wealth, fires have taken on new development trends and characteristics. Among all types of fires, building fires occur most frequently and result in the greatest losses, accounting for approximately 80% of all fires [4]. Three new types of fires—fires in large-space buildings, fires in underground transportation hubs, and fires in high-rise buildings—are occurring with increasing frequency [5,6].
With the development of socio-economy and infrastructure, the demand for large-space buildings is gradually increasing. Large-space buildings have the characteristics of good ventilation and a large amount of combustible materials inside. There are also many factors that can cause fires, such as flammable and explosive materials, daily electrical failures, and human factors. Once a fire occurs, these will further accelerate the spread of the fire. Moreover, due to the effect of hot air flow, a large-scale uncontrolled fire will gradually form.
Currently, early fire detection in large spaces primarily relies on composite detection using conventional detectors and conventional video surveillance. However, conventional detectors face numerous limitations when used for early fire detection in large spaces. For instance, when employing heat and smoke detectors, the ignition point is generally far from the detector in tall buildings. During the upward movement of smoke, it gradually disperses, making it difficult for it to accumulate and resulting in a gradual decrease in temperature. This can lead to false alarms or missed alarms. Other types of detectors also have limitations when applied to large spaces.
At present, many scholars in China have conducted extensive research in large spaces and have achieved some progress. For example, dual-mirror imaging linear beam smoke detectors and rapid flame image recognition methods have been developed.
In this paper, while studying the relevant literature on fire information detection in large spaces, we conduct research from multiple perspectives based on the work of numerous experts and scholars. After analyzing and summarizing the research content, methods, and conclusions, we grasp relevant theories and experimental methods of fire information detection in large spaces. We also share some improved schemes based on image-based fire detection algorithms and provide an outlook for future research.

2. Limitations of Conventional Fire Detection Technology in Large-Space Applications

2.1. Point-Type Smoke/Temperature-Sensing Detectors

Common Point-type Smoke/Temperature-Sensing Detectors have their core components often encapsulated in circular plastic enclosures, and are installed on building roofs or high positions on wall surfaces. For common photoelectric smoke fire detectors, the core working component is a photoelectric detector, whose working principle is shown in the figure below (Figure 1): the photoelectric sensor and the LED light source are placed on different horizontal planes. Under normal conditions, the light emitted by the LED light source does not irradiate the photoelectric sensor, thus failing to trigger the alarm mechanism. Furthermore, to further prevent the LED light from reaching the photoelectric sensor, several wedge-shaped plastic blocks are installed around the sensor to reduce the amount of light reflected from the inner wall of the enclosure that irradiates the sensor. When smoke enters the smoke detector, the large-molecule particles it contains will cause the LED light to deflect. Due to the small volume of the enclosure, some light will always reach the photoelectric sensor, thereby triggering the alarm.
Point-type Temperature-Sensing Detectors currently generally adopt a thermistor-based design (Figure 2). Each detector is equipped with two thermistors of identical performance: the reference thermistor is housed inside a shielded enclosure, while the sampling thermistor is placed externally. Due to thermal insulation of the enclosure, the reference thermistor has a slower temperature response speed; the differential temperature alarm function is achieved by utilizing the variation difference between the two thermistors. Furthermore, a fixed temperature threshold can be pre-set for the external sampling thermistor—once this threshold is reached, an alarm is triggered, thereby realizing the fixed temperature alarm function.
This detection method has limitations when applied in high and large spaces, for in such environments, buildings are tall and spacious, and the fire source is usually far from detectors. Meanwhile, smoke in the initial stage of a fire is relatively faint, and it disperses gradually as it rises, making it hard for it to accumulate [7]. Moreover, due to the large volume of the space, the temperature from the fire source gradually drops as it transmits to the sensor’s detection range, so the detector’s ability to sense abnormal temperature changes becomes less pronounced [8]. Additionally, large-space buildings are typically equipped with effective ventilation systems, which not only interfere with the operation of smoke/temperature-sensing detectors, but also may aid combustion. Thus, this method is not well-suited for fire detection in large spaces.

2.2. Line-Type Beam Smoke Fire Detectors

Common types include transmissive and reflective detectors, with the reflective type being commonly used in production and daily life [9]. Since the working principle of line-type beam smoke fire detectors (Figure 3) is to detect the weakening of light beams absorbed by fire smoke particles to trigger an alarm, in large-space environments, problems arise in the following aspects: if the installation span is too large, dust in the environment will also affect the detection results and cause false alarms; if the beam smoke detector is installed too high, the smoke generated by early fires will flow and disperse, which is not conducive to triggering the alarm of the line-type beam smoke detector; if installed too low, the detection effect will be affected by stacked goods, large equipment, artificial shielding, vibration, and other situations [10].

2.3. Infrared and Ultraviolet Flame Detectors

Infrared and ultraviolet flame detectors achieve fire monitoring by capturing characteristic optical radiation and flickering frequencies emitted by flames. During combustion, flames release electromagnetic radiation across specific wavelengths, including ultraviolet (UV) and infrared (IR) bands. These detectors typically employ ultraviolet phototubes, infrared sensors, or multi-spectrum sensors for identification.
Ultraviolet sensors utilize the photoelectric effect to generate photoelectrons and amplify signals through an avalanche discharge mechanism (Figure 4). Specifically, within a gas-filled glass tube equipped with electrodes under high voltage, incident ultraviolet radiation induces a self-sustaining discharge between the electrodes. Electrons multiply rapidly in the strong electric field, producing significant current pulses that enable effective detection of UV radiation from flames. Infrared sensors, on the other hand, distinguish flame radiation from interference caused by other heat sources using optical filtering techniques, thereby triggering alarms accurately.
Multi-spectrum sensors enhance reliability by simultaneously monitoring both UV and IR bands. This dual-channel signal verification significantly improves detection accuracy and anti-interference capability. Since flames emit ultraviolet and infrared rays of different wavelengths that are indistinguishable to the human eye, this method can detect the unique characteristics exhibited by flames in the early stages of a fire [11,12].
However, the use of infrared and ultraviolet flame detectors in large-space buildings has limitations: First, such detectors are prone to interference from other radiation sources. Common sources of interference include displays, computers, communication devices, large-scale production equipment, electronic scanning equipment, signal-transmitting devices, etc. [13]. Second, the lenses of infrared and ultraviolet fire detectors, when installed in a fixed location over time, will accumulate dust, oil stains, and other contaminants [14]. All of these factors increase the likelihood of false alarms and missed alarms for fire detection and alarm equipment.

2.4. Aspirating Smoke Fire Detectors

Aspirating smoke fire detectors (Figure 5) can actively draw in air for fire condition analysis, and generally have multiple levels of early warning and alarm prompts. Theoretically, they can quickly detect early-stage smoldering fires [15,16]. However, in large-space buildings, aspirating smoke fire detectors cannot effectively pinpoint the location of the fire source. After a low-level fire alarm is triggered, relevant personnel need to spend more time inspecting and searching for the fire source [17], which will affect the progress of quickly extinguishing the fire, or even miss the optimal opportunity, resulting in significant losses.
In order to intuitively present the core differences and commonalities among the four typical fire detectors discussed in the preceding content, a structured table is employed herein to briefly summarize and systematically organize their key characteristics (Table 1). The summary will focus on core performance dimensions of fire detectors—including, but not limited, to working principles, response sensitivity to different fire types (e.g., smoldering fire, open fire), applicable environmental conditions (e.g., high-temperature workshops, dusty warehouses), and inherent limitations (e.g., susceptibility to interference, high maintenance costs). This tabular presentation not only enhances the readability of comparative information, but also provides a clear reference framework for subsequent research on detector selection, optimization, or engineering application design.

2.5. Some Improvements in Conventional Fire Detection Techniques

In his previous work, Hou Xulong [18], through controlled variables of different combustion sources (smoldering cotton rope fire, smoldering shredded paper fire, n-heptane fire), constructed an FDS model to study the heat release rate and duration of fires under the action of different combustibles. Combined with the generation of indoor combustion products and the diffusion characteristics of smoke on the ceiling, he determined the optimal installation position for line-type beam smoke detectors: such detectors should be installed near the smoke diffusion layer. This height can not only enable effectively detection of the smoke characteristics of fires so that early warnings can be issued, but also avoid obstruction by crowds or large equipment.
For the use of detectors such as infrared and ultraviolet flame detectors in large-space buildings [19], to ensure rapid detection by the detectors, it is advisable to install them at lower positions (around 10 m) for smoldering fires with small amounts of smoke, and at the top of the building for open fires with large amounts of smoke. Meanwhile, considering the detection of fire sources with different smoke yields, a layered arrangement should be adopted when economic conditions permit for protection. However, to prevent smoke from wall reflux, the detectors in each layer should be arranged in a staggered manner [20].

2.6. The State of the Art in Large-Space Fire Detection Research

Based on the existing analysis and summary of fire detection technologies for large-space buildings, many experts and scholars at home and abroad have conducted research on early fire warning in large spaces using various methods.
Pei Yu [21] and others, in accordance with the practical needs of image-based fire detection algorithms, developed an image-based fire detection algorithm based on a target detection convolutional neural network. They conducted experiments to evaluate the detection performance of the developed fire detection algorithm based on the target detection convolutional neural network, and concluded that this algorithm has significant advantages in image-based fire detection and can accurately locate the fire position. However, there is still much room for improvement in recognition accuracy. No experiments with different interference items were carried out, and it was not proven that the target fire source can be accurately detected in the presence of interference, so the reliability of the system still needs to be verified.
Mojamed [22] and others, based on manual feature-based fire detection methods and the AlexNet architecture, manually selected and extracted features such as flame color features (such as the unique proportion range of red, green, and blue components of flames in the RGB color space), texture features (such as the irregular and dynamically changing texture characteristics presented at the edge of flames), and shape features (such as the roughly triangular or irregular polygonal outline of flames), and tried to achieve effective fire detection through the analysis and processing of these manually carefully selected features. Töreyin [23] and others identified whether a fire occurs in an area through temporal and spatial wavelet analysis. This method involves many heuristic thresholds.
The research results presented by the above researchers are helpful for understanding different large-space scenarios and environments, and also provide good enlightenment for further research on early fire detection in large spaces: at present, in research on fire recognition in large spaces, the problems of sensitivity and reliability need to be further solved; the use of set thresholds for recognition also has certain deficiencies in anti-interference.
The direct inspiration for our team is as follows: in research on large-space fire detection based on image-based flame detectors, it is necessary not only to realize the recognition and judgment of fires, but also to improve the sensitivity and reliability of fire recognition. That is, it is necessary to balance these two aspects, further improve the recognition accuracy of non-fire objects, and achieve a balance between the detection speed and accuracy of the algorithm.

3. Image-Based Large-Space Fire Detection Technology

3.1. Characteristics of Image-Based Large-Space Fire Detection Technology

This technology directly collects images from surveillance videos or processes the images [24]. By using dual-color imaging technology, it conducts a comprehensive analysis of multi-dimensional information such as heat, color, movement, shape, spectrum, and dynamics of scene images in the monitored area [25,26,27]. It constructs spectral feature discrimination models, stability models, stroboscopic feature models, flame structure feature models, and growth trend models, as well as flame feature extraction models suitable for high-space environments, so as to realize early detection and positioning of fires in tall and large spaces. Imaging flame detectors have a large protection area, a long detection distance, a fast response speed, good stability, and high reliability. They can realize the integrated application of fire prevention, anti-theft, and monitoring, showing more superior characteristics than other detectors [28,29].

3.2. Improvement Measures

Based on the summary of the above research review, although image-based fire detection technology has advantages and characteristics in fire detection in large-space environments, there are still many issues that need to be studied: (1) in-depth analysis and research on detection algorithms adapted to different fire situations; (2) continuous development and improvement of new fire detectors based on new fire detection principles; (3) comprehensive application of laser detection technology, computer vision technology, signal processing technology, and digital communication technology; (4) composite non-contact fire detection technology.
The research focuses on large-space buildings and studies fire detection methods based on image processing and temperature measurement, aiming to quickly and accurately identify whether a fire occurs, and provide new ideas for large-space fire detection (Figure 6). Combining image detection with modern network transmission technology can realize real-time monitoring and remote transmission of information, break the limitations of environment, region and distance, truly achieve synchronous sharing of remote real-time information resources and intelligent identification and analysis, and thus efficiently protect people’s lives and property. It is of great significance to society in its capability to promote more rapid and accurate fire detection.

3.3. Target Recognition Algorithm

Based on research on large-space fire detection technology using image processing, the current mainstream target recognition algorithms are as shown in the following table (Table 2).
One type is the Two-Stage algorithm represented by Faster R-CNN. Its target detection mainly consists of two parts: (1) candidate boxes are generated through a dedicated module to find the foreground and adjust the bounding boxes; (2) the region proposals (RoI, Region of Interest) generated in the first stage undergo refined detection. The other type is the One-Stage algorithm represented by SSD and YOLO, which directly performs classification and adjusts bounding boxes based on anchors [47].
Each of the two methods has its own characteristics. The Two-Stage algorithm obviously has a higher detection accuracy but a slower detection speed; the One-Stage algorithm sacrifices high accuracy, but gains speed, and is much faster than the Two-Stage algorithm.
The following is a brief introduction to the characteristics of several common Two-Stage and One-Stage algorithms: (1) Compared with the Faster-RCNN [48] algorithm, in order to reduce the amount of calculation, the R-FCN [49] algorithm moves the ROI (Region of Interest) pooling operation to before the output layer, performs ROI pooling on the feature map generated by the last layer of the network, and integrates the target’s position information into ROI pooling through position-sensitive score maps, improving the sensitivity of the feature map to positions. (2) Since both the Faster-RCNN algorithm and the R-FCN algorithm have a separate parallel candidate region extraction network in the convolutional neural network, there is a problem of slow detection speed. (3) The SSD [50] algorithm is composed of a feed-forward convolutional network and belongs to an end-to-end network architecture without region proposal, which solves the problem of slow detection speed. However, because the SSD algorithm selects a shallower network for detection, resulting in insufficient feature extraction and inability to distinguish targets well, the accuracy of this algorithm in detecting small targets is not high. (4) The YOLO v3 [51] algorithm improves the accuracy of target detection by drawing on the idea of residual networks on the basis of the SSD algorithm, and at the same time, its end-to-end network architecture significantly improves the detection speed and stability. In addition, the results generated by the YOLO v3 algorithm in predicting the categories of candidate regions are multi-label and multi-class, that is, one candidate box can belong to both smoke and flame, which lays a foundation for detecting regions where smoke and flame appear simultaneously. Therefore, to apply these common algorithms to large-space fire detection, experiments are needed to select the most suitable one.
In the target detection algorithm system, a series of key detection effect evaluation indicators play a crucial role in measuring algorithm performance. Precision [52] is expressed by the formula p_r = tp/(tp + fp), where tp refers to the number of samples correctly predicted as positive examples (True Positive), and fp is the number of samples incorrectly predicted as positive examples (False Positives). Precision measures the proportion of truly positive examples among all the results predicted as correct by the model, intuitively reflecting the probability that positive samples are accurately predicted in the prediction results. Recall [24] has the formula r_rate = tp/(tp + fn), where fn represents the number of samples incorrectly predicted as negative examples (False Negatives). Recall focuses on the proportion of real positive examples successfully predicted by the model among the total number of actually existing positive samples, and it can effectively reflect the missed-detection situation of the model, that is, how many positive samples that should have been detected are missed. The false detection rate is calculated as f_r = fp/(fp + tn), where tn is the number of samples correctly predicted as negative examples (True Negatives). The false detection rate reflects the proportion of samples that are actually incorrectly judged as negative examples (should be positive examples) among all samples judged as negative examples by the model, and is used to measure the probability that the model mistakenly identifies positive samples as negative samples. The missed detection rate is obtained by fn/(fn + tp), which directly shows the proportion of actually positive samples that are mistakenly judged as negative samples by the model, clearly demonstrating the degree of missed detection of positive samples by the model. These indicators comprehensively and meticulously depict the performance of the target detection algorithm in predicting positive and negative examples from different dimensions, providing a strong basis for evaluating algorithm performance, and helping algorithm developers to continuously optimize the model and improve detection effects.
In the field of fire detection in large-space buildings, one of the core tasks is to develop a fire detection algorithm model that is suitable for large-space buildings, with the aim of significantly improving the speed and accuracy of fire alarms in large spaces. In this regard, the target recognition algorithm plays a key role, and its basic process is as follows (Figure 7): First, collect a large amount of fire image data, use data augmentation technology to expand data diversity, carefully perform data annotation with the tool Labelimg [53], and then use the annotated data to train the convolutional neural network for fire recognition, so as to obtain the convolutional neural network model for fire recognition and the corresponding weight files. Next, import the pre-trained convolutional neural network model, import the corresponding weight files synchronously, and set relevant parameters reasonably. After completing the above preparations, read the collected images or videos, use the built and configured model to identify the images, and judge whether a fire has occurred. After going through this series of steps, accurate fire recognition is finally achieved.
To improve detection accuracy and speed, it is necessary to modify the original algorithms. The main aspects are as follows. To significantly enhance the accuracy and speed of fire detection in large-space buildings, comprehensive optimization of the original algorithms is urgently required. At the backbone network level, lightweight transformation is performed, and the attention mechanism is integrated to enable the model to focus more computing power on key features. Meanwhile, structural optimization is carried out for the neck network, the head network is adjusted reasonably, or improvements are made to the loss function and non-maximum suppression (nms) process to improve the overall model performance.
Adding a small target detection layer is also one of the key directions. By splicing shallow feature maps and deep feature maps, where shallow feature maps contain rich positional details and deep feature maps have strong semantic information, the detection effect on small-sized fire targets can be greatly improved after the fusion of the two for detection.
In the head network part, a semantic segmentation detection head is innovatively added, allowing the model to not only perform target detection and accurately locate the fire position, but also conduct semantic segmentation of the fire scene, identify different object categories and scene regions, and provide more abundant information for subsequent analysis.
Introducing a global scheduling mechanism is also crucial. Through in-depth optimization of the attention mechanism, the loss of information during transmission is reduced, and the interactive expression of global information is strengthened, so that the performance of the deep neural network is significantly improved when processing complex fire scene data, further improving detection accuracy.
In addition, during the network construction and optimization process, various convolution modules are actively tried, such as depthwise separable convolution and dilated convolution. At the same time, loss functions are boldly selected that are more suitable for fire detection scenarios, such as Distance Intersection over Union (diou) [54], Generalized Intersection over Union (giou) [55], and Focaler Intersection over Union (Focaler-iou) [56], and classic network architectures, such as ResNeSt [57], densenet [58], and resnet [59], are also introduced to draw on their strengths. Moreover, using various improved pyramid pooling technologies [60,61], such as spatial pyramid pooling [62] and adaptive pyramid pooling [63], enhances the model’s perception ability for multi-scale fire targets and comprehensively improves the performance of the large-space fire detection algorithm.

4. Innovation Directions in the Application of Image-Based Fire Detection Technology in Large-Space Environments

4.1. Selection of Image Acquisition Equipment and Fire Detection Equipment

Image-based fire detection technology utilizes visible light or surveillance cameras for fire recognition, and is particularly applicable to indoor environments (such as offices, schools, hospitals, etc.), outdoor environments (such as urban streets, industrial parks, warehouses, etc.), and transportation hubs (such as airports, stations, ports, etc.), as well as special places like museums and libraries. These scenarios typically have good lighting conditions or are widely equipped with surveillance cameras, enabling rapid capture of flame and smoke features through image recognition technology, thus achieving fire early warning with high sensitivity, real-time monitoring, and strong adaptability, and effectively improving the accuracy and timeliness of fire early warning.
Infrared cameras, in image-based fire detection, are especially suitable for scenarios such as heavy smoke environments, nighttime or low-light conditions, detection of hidden fire sources, industrial sites, and forest fire monitoring, as well as post-disaster assessment and hidden danger investigation. They can penetrate smoke, are not restricted by lighting conditions, and can quickly detect abnormal high temperatures and hidden fires by capturing thermal signals, thereby realizing highly sensitive fire early warning.
Cameras for binocular recognition in image fire detection mainly include binocular multispectral cameras combining visible light and near-infrared [64,65], and binocular CCD cameras [66,67]. These binocular cameras can achieve accurate recognition and positioning of fires by collecting images of different spectra or temperatures and combining them with intelligent algorithms. Binocular recognition technology is applicable to various scenarios, and performs particularly excellently in large and tall spaces (such as airports, squares), complex background environments (such as mines), long-distance detection scenarios, places with high requirements for positioning accuracy (industrial workshops, warehouses), and crowded places requiring real-time monitoring and early warning (such as shopping malls and exhibition halls). In these scenarios, binocular recognition systems can effectively overcome the limitations of monocular cameras, provide more accurate fire detection and positioning, and thus improve the efficiency of fire early warning and emergency response.

4.2. Construction of the Dataset

The quality of construction of a fire detection dataset directly determines the detection performance of subsequent algorithm models. Its core lies in systematically integrating the foundational nature of sample quantity, the richness of scene coverage, and adaptability to application scenarios, supplemented by high-precision annotation information (Figure 8). In terms of sample quantity, the goal is not simply to pursue the accumulation of scale, but rather to focus on the representativeness and coverage of samples. It is necessary to systematically include characteristic samples throughout the entire life cycle of fire evolution—from faint sparks and diffused smoke in the initial stage, to local open flames and concentrated smoke in the development stage, and then to large-scale flames and rolling thick smoke in the violent stage.
In Andrean D et al.’s study [68], the dataset centered on “fire (hotspots) and smoke” as core detection targets, comprising 746 original images covering two scenarios: forest fires and candle flames. Post-augmentation, it was subdivided into three targeted datasets (1341 forest fire-only, 608 candle flame-only, and 1790 mixed-scenario images) to verify the model’s capability in detecting fires of varying scales. The dataset processing was both standardized and practical: annotations were conducted via app.roboflow.com, with all original images labeled using bounding boxes for target framing and “fire”/“smoke” category tags to focus training on core features. Data augmentation employed four techniques—saturation adjustment (±50%), 25% cropping, 20-degree shear deformation, and 90-degree multi-directional rotation—generating 2–4 new images per original without additional data collection to enhance diversity. Additionally, scenario-based dataset subdivision enabled precise alignment with different testing objectives for model performance validation. This ensured that the model could learn the morphological characteristics and dynamic change laws of different combustion stages.
The construction of scene richness needs to break through the limitations of a single environment and comprehensively cover multi-dimensional interference factors: lighting conditions should include complex situations such as direct natural light, dim light, backlight, and glare; meteorological environments need to cover different weather states such as sunny, rainy, foggy and dusty; spatial scenes should span various places, including indoor (such as kitchens, warehouses, residential buildings) and outdoor (such as forests, industrial parks, construction sites). At the same time, it is necessary to specifically include samples of easily confused targets (such as candle flames, kitchen fumes, industrial steam, dust clusters, etc.) to enhance the model’s ability to distinguish between fire and non-fire categories [69,70].
In the study by El-Madafri et al. [71], an RGB forest fire dataset named “Wildfire Dataset” was constructed, containing 2700 images. The dataset was divided into two main categories (“fire” and “nofire”) and five subcategories (including fire smoke and confounding elements), covering diverse environmental conditions and confounding factors. The data were sourced from public channels such as government agencies and Flickr, all of which are public domain-licensed, and a CSV file was attached to record each image’s URL and resolution for traceability. In terms of dataset processing, after perceptual hash-based deduplication, the dataset was split into 70% for training, 15% for validation, and 15% for testing, with random shuffling. To address the imbalance between the “fire” class (1047 images) and “nofire” class (1653 images), only the “fire” subcategories in the training set were augmented to match the size of the “nofire” class, while the validation and test sets retained their natural distribution. Additionally, the impact of the weights of confounding element subcategories on model training was analyzed to enhance the model’s ability to distinguish confounding factors.
To tackle the high false alarm and missed alarm rates of single-spectrum image-based forest fire monitoring, Liu et al. [72] built a multispectral (visible + infrared) dataset consisting of two parts: a scenario part and a fire part. The scenario part contained 6972 image pairs (one visible image and one infrared image per pair), covering multiple elements (e.g., humans, roads, buildings), as well as day/night and different altitude scenarios. The fire part included 7193 images (simulating forest fires using dry branches and leaves as fuel) covering categories such as smoke, fire, and ash. The data were acquired by synchronously capturing videos with a DJI UAV and extracting frames (one frame every five frames). During preprocessing, visible images were cropped to 512 × 512 pixels to align with infrared images. The images were then fused via FF-Net (incorporating the CBAM attention mechanism), with optimization using a combined loss function of M-SSIM and TV. Finally, YOLOv5 was employed for detection. The fused images exhibited significantly lower false alarm and missed alarm rates than single-spectrum images, improving the reliability of fire detection.
Scene adaptability is key to improving the practical value of the dataset, and it is necessary to carry out customized design in close combination with the characteristics of specific application scenarios [73,74,75]. For example, a dataset for forest fire monitoring should focus on collecting samples of early smoke diffusion and surface fire spread against the background of dead branches and fallen leaves; for fire detection in indoor scenarios, it is necessary to increase the proportion of interference samples such as kitchen fumes, lighter flames, and heat sources of heating equipment to reduce the false alarm rate in practical applications.
In addition, precision control of the annotation link is also indispensable. It is necessary to accurately classify and mark the bounding boxes of targets such as open flames and smoke to avoid model learning deviations caused by ambiguous or ambiguous annotations. Only through the collaborative optimization of the above multiple dimensions can the constructed dataset provide high-quality training samples for the fire detection model, thereby effectively improving the detection accuracy and robustness of the model in complex environments.

4.3. Algorithm Improvement

In recent years, numerous studies have focused on improving object detection algorithms for fire detection. However, it is not always the case that more advanced algorithms yield better results. Instead, the key lies in the applicability and adaptability of the algorithms [76]. Only by selecting appropriate algorithms and making targeted improvements can the best results be achieved.
Chong Wang [77] et al. specifically improved the algorithm for the smoke recognition task by proposing a Cross-Contrastive Patch Embedding (CCPE) module based on the Swin Transformer. This module enhances the network’s ability to distinguish low-level details by leveraging multi-scale spatial contrast information in both vertical and horizontal directions. The integration of cross-contrast with the Transformer not only capitalizes on the Transformer’s strengths in global receptive fields and context modeling, but also compensates for its limitations in capturing low-level details. The researchers introduced the Separable Negative Sampling Mechanism (SNSM) to address the issue of supervision signal confusion during training and utilized the SKLFS-WildFire test dataset, which is the largest real-world wildfire test set to date. The method showed significant performance improvements compared to the original detection model. In terms of performance, the image-level AUC increased from 0.765 to 0.900 (+13.5%), serving as a core indicator for wildfire smoke identification; the video-level AUC rose from 0.908 to 0.934 (+2.6%), adapting to real-world continuous frame monitoring scenarios; and the bounding box AP@0.1 improved from 0.476 (Swin-Tiny baseline) to 0.537 (+6.1%), addressing the Transformer’s deficiency in capturing low-level details. Additionally, in the multi-frame task on the FIgLib dataset, the algorithm achieved optimal performance in terms of accuracy (Acc), F1-score, and recall, with the detection time reduced by 23%. In terms of efficiency, YOLOX-ContrastSwinT has a parameter count of only 35.893 million (lower than that of the baseline YOLOX and Sparse R-CNN) and a GFLOPs value of 53.250 (merely one third of that of the baseline YOLOX). It supports real-time deployment on embedded devices, with a latency of 116 ms per frame on the RK3588 platform and 89 ms per frame on the Jetson Orin Nano platform.
Furkat Safarov [78] et al. specifically improved fire detection in complex and occluded environments by proposing a novel fire and smoke detection method that combines the Vision Transformer (ViT) with the YOLOv5s object detection model. By replacing the backbone network of YOLOv5 with ViT, the model’s detection accuracy for fires and smoke in complex environments was enhanced, particularly in scenes with occlusions and large-area distributions. This method provides an effective approach for real-time fire detection in both urban and natural environments. Experimental results show that this model outperforms baseline YOLOv5 variants across key metrics, with an mAP@0.5 of 0.664 and recall of 0.657.
Rui Li [79] et al. noted that smoke is an early sign of forest fires, with accurate identification critical for fire prevention and control. However, current detection faces three key issues: complex smoke texture causing detection omissions; smoke-like objects in forests interfering with recognition; and thin edge smoke leading to edge omissions. To solve these, the authors proposed a high-precision edge-focused forest fire smoke detection network, improved from original models—Swin Transformer (feature extraction backbone) and AugFPN (feature pyramid network). Specific improvements included the following: (1) the Swin Multidimensional Window Extractor (SMWE), enhancing horizontal-vertical inter-window information exchange to extract global texture features of smoke images and mitigate omissions; (2) the Guillotine Feature Pyramid Network (GFPN) with a new guillotine convolution, reducing redundant features via feature fusion to boost anti-interference ability; (3) a contour-adaptive loss function (addressing thin, irregular edge smoke) to reduce boundary blur from feature map downsampling. Experimental results showed that the SMWE-GFPNNet model achieved a mean Average Precision (mAP) of 80.92%, mAP50 of 90.01%, and mAP75 of 83.38% on the Forest Fire Smoke Complex Background Detection Dataset, with excellent anti-interference performance and accuracy.
Sixu Pu [80] et al. proposed an advanced YOLOv8-based flame segmentation scheme—trained and optimized on an independently constructed custom dataset—to address two key issues: the critical role of effective flame region segmentation in coal-fired power plant boiler burners (for enhanced combustion monitoring and operational safety), and the vulnerability of traditional methods to complex furnace interferences (e.g., furnace background, adjacent burners, fly ash). Their core improvements included the following: introducing a multi-level augment head (MAH) and single-object tracking for effective integration of shallow/deep features; embedding a convolutional block attention module (CBAM) in the backbone network to refine flame feature extraction; and adopting reparameterization and cross-stage feature reuse in the neck network to boost multi-scale feature fusion. The improved model achieved a mean intersection over union (mIoU) of 82.6% (an 8.8-unit improvement over the original YOLOv8), while retaining high computational efficiency with a 4.43 ms inference time. It outperformed other semantic segmentation models in both accuracy and speed, with practicality validated on a 660 MW opposed-fired boiler (resists external disturbances and enables accurate flame segmentation under variable loads). This work offers reliable support for image-based flame detection systems, enabling real-time, precise combustion monitoring.

4.4. Importance in the Selection of Technical Directions

Deng Li [81] et al. collected a dataset using visible-light cameras in a real aircraft hangar and subsequently employed it for fire detection. The selection of imaging equipment is critically dependent on the specific application scenario. For large-space structures like aircraft hangars, key considerations include the following: (1) Topology: high ceilings and vast spans necessitate detection systems with long-range capabilities; (2) Environmental interference: varying lighting conditions (e.g., strong sunlight through windows vs. poorly lit areas) and potential visual obstructions challenge conventional visible-light cameras; (3) Real-time requirements: the need for early warning is urgent due to the presence of highly flammable materials (e.g., aviation fuel, hydraulic fluids, and polymers) and extremely valuable assets (e.g., aircraft and machinery). Furthermore, the construction of a representative dataset is equally vital. Both the sample size and scene diversity—including different fire sources (liquid fuel fires, solid material fires) and varying illumination (day/night, direct light/shadow)—directly impact the generalization and robustness of the detection algorithm. After improving the YOLOv8n algorithm, the fire detection system demonstrated excellent performance in recognizing simulated aircraft hangar fires. The enhanced detection algorithm was effectively applied to identify fire accidents, accurately and swiftly detecting smoke and flames. Notably, it did not produce false positives or false negatives in strongly sunlit backgrounds, exhibiting strong anti-interference capabilities. However, it is worth mentioning that when using the Grad-CAM visualization technique [82,83] to analyze the network’s focus areas, it was found that the improved algorithm had relatively dispersed attention points (highlighted regions) when detecting fire smoke in darker environments (Figure 9). This suggests that the system may have certain limitations when detecting fire smoke.
Therefore, in follow-up experiments for aircraft hangar fire detection, infrared cameras will be employed for data collection and subsequent detection experiments, abandoning smoke detection in favor of achieving longer detection ranges and greater resistance to light interference in image-based fire detection systems (Figure 10). Compared with the original experimental scheme and results, a better solution has been obtained [82].
When using image-based fire detectors for fire detection, it is not only necessary to make targeted improvements to the target detection algorithm according to actual needs, but also to choose an experimental scheme that fits the actual situation, and then select the appropriate camera equipment to collect the image set required to train the network (Table 3). Using similar camera equipment in subsequent actual detection will yield better results.

5. Conclusions and Future Work

With economic development, large-space buildings are becoming increasingly common. Due to the unique characteristics of these buildings, it is essential to implement specific fire prevention measures. Large-space buildings have ample space and considerable height, resulting in a large internal volume. Conventional fire detectors are insufficient for detecting fire information in high buildings, often leading to high rates of missed and false alarms. In contrast, image-based fire detectors can effectively address these challenges, compensating for the shortcomings of conventional detectors, such as missed detections, false alarms, and untimely responses.
Practical research in aircraft hangar fire detection has shown that using an image-based fire detection system with advantages like large coverage, high sensitivity, high precision, and a short response time can effectively reduce the common issues of missed and false alarms and failure to detect fires in time. Given its unparalleled advantages over conventional detectors, the image-based method can effectively compensate for the drawbacks of conventional detectors. Therefore, it is likely to be widely applied in fire detection and alarm systems for large-space buildings. It can accurately and quickly detect fire information, reducing fire-related losses and ensuring the safe operation of businesses.
When designing fire detection systems and their improvements for large spaces, it is crucial to base the design on real-world application scenarios. Future research should prioritize several key operational directions: (1) developing advanced algorithms to reduce false alarms in complex environments with dynamic lighting and industrial interferents; (2) establishing comprehensive, multi-scenario fire image datasets that encompass diverse large-space settings to support robust model training; (3) exploring optimized multi-modal data fusion techniques that intelligently combine information from various sensors, including visible-light, thermal infrared, and binocular cameras, to enhance detection reliability; and (4) investigating cost-effective hardware solutions and edge-computing architectures to address the practical challenges of high computational resource demands and system deployment costs. Careful experimentation and in-depth research are needed to determine the appropriate equipment and methods for algorithm improvement to achieve the best detection results.

Author Contributions

Conceptualization, L.D. and Q.L.; literature search, S.W. and S.Z.; literature analysis and synthesis, L.D., S.W. and S.Z.; writing—original draft preparation, S.W.; writing—review and editing, L.D., S.W. and Q.L.; visualization, S.W.; supervision, Q.L.; project administration, L.D. and Q.L.; funding acquisition, Q.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the Professor Wan Ki Chow research initiation project (No. XYKY2025002); the Civil Aviation Joint Research Fund of National Natural Science Foundation of China (U2033206); the Key Laboratory Project of Sichuan Province, No. MZ2022JB01; and the Aeronautical Science Foundation of China (ASFC-20200046117001).

Institutional Review Board Statement

Not applicable.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author.

Acknowledgments

The authors wish to thank W.K. Chow for his advice.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. CECS 263:2009; Technical Specification for Large-Space Intelligent Active Control Sprinkler Systems. China Association for Engineering Construction Standardization: Beijing, China, 2009.
  2. Hu, Y. Research on Image-based Fire Risk Recognition in Complex Large Space. Ph.D. Thesis, Xi’an University of Ar-chitecture and Technology, Xi’an, China, 2016. (In Chinese). Available online: https://kns.cnki.net/kcms2/article/abstract?v=RPbSoBw3VsFkk5qoQBYt9eWg3pC5HgNvnYaE_Da8WUjdDt8PTL8JKra53ij2ReXSjW92WXBKrtlcJcCKlX_olWk0QGGmAl06BJB6-KtEx_SdCyXeVxz6qR-QmY3sLKdVFzf4f-thB1KC9krc-urrcRUaPt_0VMneT6ewTYm8qZ1020UKPde_gtLuFyq83q9x&uniplatform=NZKPT&language=CHS (accessed on 20 June 2025).
  3. Chen, T.; Yuan, H.; Su, G.; Fan, W. An automatic fire searching and suppression system for large spaces. Fire Saf. J. 2004, 39, 297–307. [Google Scholar] [CrossRef]
  4. Jamali, M.; Samavi, S.; Nejati, M.; Mirmahboub, B. Outdoor Fire Detection based on Colorand Motion Characteristics. In Proceeding of the 21st Iranian Conference on Electrical Engineering (ICEE), Mashhad, Iran, 14–16 May 2013. [Google Scholar]
  5. Li, C.; Chen, J.; Li, H.; Hu, L.; Cao, J. Experimental research on fire spreading anddetection method of underground utility pipetunnel. Fire Sci. Technol. 2019, 38, 1258–1261. (In Chinese) [Google Scholar]
  6. Li, F.; Li, H. Fire Safety Strategies for Typical Space of Large Transportation Hubs. In Fire Protection Engineering Applications for Large Transportation Systems in China; Springer International Publishing: Cham, Switzerland, 2020; pp. 99–131. [Google Scholar]
  7. Cui, B.; Wang, C.; Wu, M.; Zhu, C.; Wang, D.; Li, B. Integrating Bluetooth-enabled sensors with cloud computing for fire hazard communication systems. ASCE-ASME J. Risk Uncertain. Eng. Syst. Part A Civ. Eng. 2024, 10, 04024035. [Google Scholar] [CrossRef]
  8. Gaur, A.; Singh, A.; Kumar, A.; Kulkarni, K.S.; Lala, S.; Kapoor, K. Fire sensing technologies: A review. IEEE Sens. J. 2019, 19, 3191–3202. [Google Scholar] [CrossRef]
  9. Yao, Y.; Yuyu, T.; Peng, C. Study on calibration method of filter for linear beam smoke detector AOPC 2022: Optoelec-tronics and Nanophotonics. SPIE 2023, 12556, 114–119. [Google Scholar]
  10. Meacham, B.J. Factors affecting the early detection of fire in electronic equipment and cable installations. Fire Technol. 1993, 29, 34–59. [Google Scholar] [CrossRef]
  11. Long, B. Research and Design of the Comprehensive Test System Based on Ultraviolet Flame Detector. Ph.D. Thesis, Xi’an Polytechnic University, Xi’an, China, 2013. (In Chinese). Available online: https://kns.cnki.net/kcms2/article/abstract?v=RPbSoBw3VsFL_R6t2Q-HsNTeiFZeS0bfmgnDWinJ2QMMAoClJqeMdblT6nOuqWMAarb67hcbsMcrjHR6HtcSBWT0o8ZxYiqK5mBL9akcmLgeEUNgu7xz0Ld1aZx8qCKBB9h8PnQAVAuZ0cjERTuOvHHFmrp_zS_q2lw_ZlS_vayc14CAFYMSsfVkmUXLSDEB&uniplatform=NZKPT&language=CHS (accessed on 20 June 2025).
  12. Tsai, C.F.; Young, M.S. Measurement system using ultraviolet and multiband infrared technology for identifying fire behavior. Rev. Sci. Instrum. 2006, 77, 014901. [Google Scholar] [CrossRef]
  13. Bordbar, H.; Alinejad, F.; Conley, K.; Ala-Nissila, T.; Hostikka, S. Flame detection by heat from the infrared spectrum: Optimization and sensitivity analysis. Fire Saf. J. 2022, 133, 103673. [Google Scholar] [CrossRef]
  14. Pauchard, A.R.; Manic, D.; Flanagan, A.; Besse, P.A.; Popovic, R.S. A method for spark rejection in ultraviolet flame detectors. IEEE Trans. Ind. Electron. 2000, 47, 168–174. [Google Scholar] [CrossRef]
  15. Li, Q.; Yue, L. Research on the Design Improvement of Aspirating Smoke Fire Detector. Fire Sci. Technol. 2021, 40, 1644–1647. (In Chinese) [Google Scholar]
  16. Lee, Y.M.; Khieu, H.T.; Kim, D.W.; Kim, J.T.; Ryou, H.S. Predicting the Fire Source Location by Using the Pipe Hole Network in Aspirating Smoke Detection System. Appl. Sci. 2022, 12, 2801. [Google Scholar] [CrossRef]
  17. Višak, T.; Baleta, J.; Virag, Z.; Vujanović, M.; Wang, J.; Qi, F. Multi objective optimization of aspirating smoke detector sampling pipeline. Optim. Eng. 2021, 22, 121–140. [Google Scholar] [CrossRef]
  18. Hou, X.L.; Jin, W.J. Numerical Simulation of Fires and Analysis of Detection Response in Large-scale Building Spaces. Technol. Wind. 2014, 164–165+167. [Google Scholar] [CrossRef]
  19. Xu, F.; Zhang, X. Test on application of flame detector for large space environment. Procedia Eng. 2013, 52, 489–494. [Google Scholar] [CrossRef]
  20. Khan, F.; Xu, Z.; Sun, J.; Khan, F.M.; Ahmed, A.; Zhao, Y. Recent advances in sensors for fire detection. Sensors 2022, 22, 3310. [Google Scholar] [CrossRef]
  21. Yu, P.; Wei, W.; Li, J.; Du, Q.; Wang, F.; Zhang, L.; Li, H.; Yang, K.; Yang, X.; Zhang, N.; et al. Fire-PPYOLOE: An Efficient Forest Fire Detector for Real-Time Wild Forest Fire Monitoring. J. Sens. 2024, 2024, 2831905. [Google Scholar] [CrossRef]
  22. Al Mojamed, M. Smart Mina: LoRaWAN Technology for Smart Fire Detection Application for Hajj Pilgrimage. Comput. Syst. Sci. Eng. 2022, 40, 259. [Google Scholar] [CrossRef]
  23. Töreyin, B.U.; Dedeoğlu, Y.; Güdükbay, U.; Çetin, A.E. Computer vision based method for real-time fire and flame detection. Pattern Recognit. Lett. 2006, 27, 49–58. [Google Scholar] [CrossRef]
  24. Chitram, S.; Kumar, S.; Thenmalar, S. Enhancing Fire and Smoke Detection Using Deep Learning Techniques. Eng. Proc. 2024, 62, 7. [Google Scholar] [CrossRef]
  25. Kim, Y.H.; Kim, A.; Jeong, H.Y. RGB color model based the fire detection algorithm in video sequences on wireless sensor network. Int. J. Distrib. Sens. Netw. 2014, 10, 923609. [Google Scholar] [CrossRef]
  26. Khalil, A.; Rahman, S.U.; Alam, F.; Ahmad, I.; Khalil, I. Fire detection using multi color space and background modeling. Fire Technol. 2021, 57, 1221–1239. [Google Scholar] [CrossRef]
  27. Yeh, C.H.; Liu, Y.H. Development of Two-Color pyrometry for flame impingement on oxidized metal surfaces. Exp. Therm. Fluid Sci. 2024, 152, 111108. [Google Scholar] [CrossRef]
  28. Hou, J.; Qian, J.; Zhang, W.; Zhao, Z.; Pan, P. Fire detection algorithms for video images of large space structures. Multimed. Tools Appl. 2011, 52, 45–63. [Google Scholar] [CrossRef]
  29. Zhang, G.; Cui, R.; Qi, K.; Wang, B. Research on Large Space Fire Monitoring Based on Image Processing. J. Phys. 2021, 2074, 012003. [Google Scholar] [CrossRef]
  30. Xie, X.; Cheng, G.; Wang, J.; Yao, X.; Han, J. Oriented R-CNN for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Montreal, QC, Canada, 10–17 October 2021; pp. 3520–3529. [Google Scholar]
  31. Lu, X.; Li, B.; Yue, Y.; Li, Q.; Yan, J. Grid r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 15–20 June 2019; pp. 7363–7372. [Google Scholar]
  32. Redmon, J.; Farhadi, A. Yolov3: An incremental improvement. arXiv 2018, arXiv:1804.02767. [Google Scholar] [CrossRef]
  33. Zheng, H.; Duan, J.; Dong, Y.; Liu, Y. Real-time fire detection algorithms running on small embedded devices based on MobileNetV3 and YOLOv4. Fire Ecol. 2023, 19, 31. [Google Scholar] [CrossRef]
  34. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2026; pp. 779–788. [Google Scholar]
  35. Terven, J.; Córdova-Esparza, D.M.; Romero-González, J.A. A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas. Mach. Learn. Knowl. Extr. 2023, 5, 1680–1716. [Google Scholar] [CrossRef]
  36. Zhang, Y.; Jiao, Y.; Dou, Y.; Zhao, L.; Liu, Q.; Zuo, G. A Lightweight Dynamically Enhanced Network for Wildfire Smoke Detection in Transmission Line Channels. Processes 2025, 13, 349. [Google Scholar] [CrossRef]
  37. Khanam, R.; Hussain, M. Yolov11: An overview of the key architectural enhancements. arXiv 2024, arXiv:2410.17725. [Google Scholar] [CrossRef]
  38. Tian, Y.; Ye, Q.; Doermann, D. Yolov12: Attention-centric real-time object detectors. arXiv 2025, arXiv:2502.12524. [Google Scholar] [CrossRef]
  39. Girshick, R. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, Washington, DC, USA, 7–13 December 2015; pp. 1440–1448. [Google Scholar]
  40. Cheng, X.; Yu, J. RetinaNet with difference channel attention and adaptively spatial feature fusion for steel surface defect detection. IEEE Trans. Instrum. Meas. 2020, 70, 1–11. [Google Scholar] [CrossRef]
  41. Cai, J.; Zhang, L.; Dong, J.; Guo, J.; Wang, Y.; Liao, M. Automatic identification of active landslides over wide areas from time-series InSAR measurements using Faster RCNN. Int. J. Appl. Earth Obs. Geoinf. 2023, 124, 103516. [Google Scholar] [CrossRef]
  42. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 2016, 39, 1137–1149. [Google Scholar] [CrossRef]
  43. He, K.; Gkioxari, G.; Dollar, P.; Girshick, R. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision 2017, Venice, Italy, 22–29 October 2017; pp. 2961–2969. [Google Scholar]
  44. He, K.; Zhang, X.; Ren, S.; Sun, J. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans. Pattern Anal. Mach. Intell. 2015, 37, 1904–1916. [Google Scholar] [CrossRef]
  45. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.-Y.; Berg, A.C. Ssd: Single shot multibox detector. In Proceedings of the Computer Vision–ECCV 2016: 14th European Conference (Proceedings, Part I 14), Amsterdam, The Netherlands, 11–14 October 2016; Springer International Publishing: Berlin/Heidelberg, Germany, 2016; pp. 21–37. [Google Scholar]
  46. Xu, G.; Zhang, Q.; Liu, D.; Lin, G.; Wang, J.; Zhang, Y. Adversarial adaptation from synthesis to reality in fast detector for smoke detection. IEEE Access 2019, 7, 29471–29483. [Google Scholar] [CrossRef]
  47. Padilla, R.; Netto, S.L.; da Silva, E.A.B. A Survey on Performance Metrics for Object-Detection Algorithms. In Proceedings of the 2020 International Conference on Systems, Signals and Image Processing (IWSSIP), Niteroi, Brazil, 1–3 July 2020. [Google Scholar]
  48. Liu, Y.; Wu, D.; Liang, J.; Wang, H. Aeroengine blade surface defect detection system based on improved faster RCNN. Int. J. Intell. Syst. 2023, 2023, 1992415. [Google Scholar] [CrossRef]
  49. Zhang, Y.; Chi, M. Mask-R-FCN: A deep fusion network for semantic segmentation. IEEE Access 2020, 8, 155753–155765. [Google Scholar] [CrossRef]
  50. Zhao, H.; Gao, Y.; Deng, W. Defect detection using shuffle Net-CA-SSD lightweight network for turbine blades in IoT. IEEE Internet Things J. 2024, 11, 32804–32812. [Google Scholar] [CrossRef]
  51. Li, P.; Zhao, W. Image fire detection algorithms based on convolutional neural networks. Case Stud. Therm. Eng. 2020, 19, 100625. [Google Scholar] [CrossRef]
  52. Luan, T.; Zhou, S.; Liu, L.; Pan, W. Tiny-object detection based on optimized YOLO-CSQ for accurate drone detection in wildfire scenarios. Drones 2024, 8, 454. [Google Scholar] [CrossRef]
  53. Lin, T. LabelImg, 2015. Available online: https://github.com/tzutalin/labelImg (accessed on 28 June 2025).
  54. Zheng, Z.; Wang, P.; Liu, W.; Li, J.; Ye, R.; Ren, D. Distance-IoU loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI Conference on Artificial Intelligence, Philadelphia, PA, USA, 25February–4 March 2025; Volume 34, pp. 12993–13000. [Google Scholar]
  55. Rezatofighi, H.; Tsoi, N.; Gwak, J.; Sadeghian, A.; Reid, I.; Savarese, S. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 15–20 June 2019; pp. 658–666. [Google Scholar]
  56. Zhang, H.; Zhang, S. Focaler-iou: More focused intersection over union loss. arXiv 2024, arXiv:2401.10525. [Google Scholar] [CrossRef]
  57. Lian, Q.; Luo, X.; Lin, D.; Lin, C.; Chen, B.; Guo, Z. ResNest-SVM-based method for identifying single-phase ground faults in active distribution networks. Front. Energy Res. 2024, 12, 1501737. [Google Scholar] [CrossRef]
  58. Siddiqui, F.; Yang, J.; Xiao, S.; Fahad, M. Enhanced deepfake detection with DenseNet and Cross-ViT. Expert Syst. Appl. 2025, 267, 126150. [Google Scholar] [CrossRef]
  59. Razavi, M.; Mavaddati, S.; Koohi, H. ResNet deep models and transfer learning technique for classification and quality detection of rice cultivars. Expert Syst. Appl. 2024, 247, 123276. [Google Scholar] [CrossRef]
  60. Wang, S.; Wu, M.; Wei, X.; Song, X.; Wang, Q.; Jiang, Y.; Gao, J.; Meng, L.; Chen, Z.; Zhang, Q.; et al. An advanced multi-source data fusion method utilizing deep learning techniques for fire detection. Eng. Appl. Artif. Intell. 2025, 142, 109902. [Google Scholar] [CrossRef]
  61. Elhassan, M.A.M.; Zhou, C.; Benabid, A.; Adam, A.B.M. P2AT: Pyramid pooling axial transformer for real-time semantic segmentation. Expert Syst. Appl. 2024, 255, 124610. [Google Scholar] [CrossRef]
  62. Dewi, C.; Chen, R.-C.; Yu, H.; Jiang, X. Robust detection method for improving small traffic sign recognition based on spatial pyramid pooling. J. Ambient Intell. Humaniz. Comput. 2023, 14, 8135–8152. [Google Scholar] [CrossRef]
  63. Li, Z.; He, Q.; Zhao, H.; Yang, W. Doublem-net: Multi-scale spatial pyramid pooling-fast and multi-path adaptive feature pyramid network for UAV detection. Int. J. Mach. Learn. Cybern. 2024, 15, 5781–5805. [Google Scholar] [CrossRef]
  64. Sun, W.; Liu, Y.; Wang, F.; Hua, L.; Fu, J.; Hu, S. A Study on Flame Detection Method Combining Visible Light and Thermal Infrared Multimodal Images. Fire Technol. 2024, 61, 2167–2188. [Google Scholar] [CrossRef]
  65. Sun, F.; Yang, Y.; Lin, C.; Liu, Z.; Chi, L. Forest fire compound feature monitoring technology based on infrared and visible binocular vision. Journal of Physics: Conference Series. IOP Publ. 2021, 1792, 012022. [Google Scholar]
  66. Song, T.; Tang, B.; Zhao, M.; Deng, L. An accurate 3-D fire location method based on sub-pixel edge detection and non-parametric stereo matching. Measurement 2014, 50, 160–171. [Google Scholar] [CrossRef]
  67. Li, G.; Lu, G.; Yan, Y. Fire detection using stereoscopic imaging and image processing techniques. In Proceedings of the 2014 IEEE International Conference on Imaging Systems and Techniques (IST) Proceedings, Santorini, Greece, 14–17 October 2014; pp. 28–32. [Google Scholar]
  68. Andrean, D.; Unik, M.; Rizki, Y. Hotspots and Smoke Detection from Forest and Land Fires Using the YOLO Algorithm (You Only Look Once). JIM-J. Int. Multidiscip. 2023, 1, 46–56. [Google Scholar]
  69. Gragnaniello, D.; Greco, A.; Sansone, C.; Vento, B. Fire and smoke detection from videos: A literature review under a novel taxonomy. Expert Syst. Appl. 2024, 255, 124783. [Google Scholar] [CrossRef]
  70. Wang, M.; Yue, P.; Jiang, L.; Yu, D.; Tuo, T.; Li, J. An open flame and smoke detection dataset for deep learning in remote sensing based fire detection. Geo-Spat. Inf. Sci. 2025, 28, 511–526. [Google Scholar] [CrossRef]
  71. El-Madafri, I.; Peña, M.; Olmedo-Torre, N. The wildfire dataset: Enhancing deep learning-based forest fire detection with a diverse evolving open-source dataset focused on data representativeness and a novel multi-task learning approach. Forests 2023, 14, 1697. [Google Scholar] [CrossRef]
  72. Liu, Y.; Zheng, C.; Liu, X.; Tian, Y.; Zhang, J.; Cui, W. Forest Fire Monitoring Method Based on UAV Visual and Infrared Image Fusion. Remote Sens. 2023, 15, 3173. [Google Scholar] [CrossRef]
  73. Pincott, J.; Tien, P.W.; Wei, S.; Calautit, J.K. Indoor fire detection utilizing computer vision-based strategies. J. Build. Eng. 2022, 61, 105154. [Google Scholar] [CrossRef]
  74. Gaur, A.; Singh, A.; Kumar, A.; Kumar, A.; Kapoor, K. Video flame and smoke based fire detection algorithms: A literature review. Fire Technol. 2020, 56, 1943–1980. [Google Scholar] [CrossRef]
  75. Yar, H.; Khan, Z.A.; Ullah, F.U.M.; Ullah, W.; Baik, S.W. A modified YOLOv5 architecture for efficient fire detection in smart cities. Expert Syst. Appl. 2023, 231, 120465. [Google Scholar] [CrossRef]
  76. Casas, E.; Ramos, L.; Bendek, E.; Rivas-Echeverria, F. Yolov5 vs. yolov8: Performance benchmarking in wildfire and smoke detection scenarios. J. Image Graph. 2024, 12, 127–136. [Google Scholar] [CrossRef]
  77. Wang, C.; Xu, C.; Akram, A.; Wang, Z.; Shan, Z.; Zhang, Q. Wildfire Smoke Detection System: Model Architecture, Training Mechanism, and Dataset. Int. J. Intell. Syst. 2025, 2025, 1610145. [Google Scholar] [CrossRef]
  78. Safarov, F.; Muksimova, S.; Kamoliddin, M.; Cho, Y.I. Fire and smoke detection in complex environments. Fire 2024, 7, 389. [Google Scholar] [CrossRef]
  79. Li, R.; Hu, Y.; Li, L.; Guan, R.; Yang, R.; Zhan, J.; Cai, W.; Wang, Y.; Xu, H.; Li, L. SMWE-GFPNNet: A high-precision and robust method for forest fire smoke detection. Knowl.-Based Syst. 2024, 289, 111528. [Google Scholar] [CrossRef]
  80. Pu, S.; Li, J.; Han, Z.; Zhu, X.; Xu, C. Flame region segmentation of a coal-fired power plant boiler burner through YOLOv8-MAH model. Fuel 2025, 398, 135518. [Google Scholar] [CrossRef]
  81. Deng, L.; Wu, S.; Zhou, J.; Zou, S.; Liu, Q. LSKA-YOLOv8n-WIoU: An Enhanced YOLOv8n Method for Early Fire Detection in Airplane Hangars. Fire 2025, 8, 67. [Google Scholar] [CrossRef]
  82. Xiong, C.; Zayed, T.; Abdelkader, E.M. A novel YOLOv8-GAM-Wise-IoU model for automated detection of bridge surface cracks. Constr. Build. Mater. 2024, 414, 135025. [Google Scholar] [CrossRef]
  83. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 618–626. [Google Scholar]
  84. Deng, L.; Wang, Z.; Liu, Q. AH-YOLO: An Improved YOLOv8-Based Lightweight Model for Fire Detection in Aircraft Hangars. Fire 2025, 8, 199. [Google Scholar] [CrossRef]
Fire 08 00358 g001
Figure 1. Working principle of photoelectric smoke detector. (a) LED light source; (b) wedge-shaped plastic blocks; (c) deflected light; (d) photoelectric sensor; (e) smoke.
Figure 1. Working principle of photoelectric smoke detector. (a) LED light source; (b) wedge-shaped plastic blocks; (c) deflected light; (d) photoelectric sensor; (e) smoke.
Fire 08 00358 g001
Fire 08 00358 g002
Figure 2. Main structure of Point-type Temperature-Sensing Detector. (a) Adjusting resistor; (b) reference thermistor; (c) sampling thermistor; (d) threshold circuit; (e) bistable circuit.
Figure 2. Main structure of Point-type Temperature-Sensing Detector. (a) Adjusting resistor; (b) reference thermistor; (c) sampling thermistor; (d) threshold circuit; (e) bistable circuit.
Fire 08 00358 g002
Fire 08 00358 g003
Figure 3. Working principle of reflective-type line-type beam smoke fire detector. (a) Mounting bracket, (b) detector, (c) reflector.
Figure 3. Working principle of reflective-type line-type beam smoke fire detector. (a) Mounting bracket, (b) detector, (c) reflector.
Fire 08 00358 g003
Fire 08 00358 g004
Figure 4. Schematic diagram of detection circuit for UV flame detector. (a) Ultraviolet light; (b) anode; (c) current limiter; (d) high voltage; (e) electrons; (f) cathode.
Figure 4. Schematic diagram of detection circuit for UV flame detector. (a) Ultraviolet light; (b) anode; (c) current limiter; (d) high voltage; (e) electrons; (f) cathode.
Fire 08 00358 g004
Fire 08 00358 g005
Figure 5. Operating principle of aspirating smoke fire detector. (a) Inhaled air; (b) terminal cap; (c) sampling and analysis; (d) filter; (e) aspirating pump.
Figure 5. Operating principle of aspirating smoke fire detector. (a) Inhaled air; (b) terminal cap; (c) sampling and analysis; (d) filter; (e) aspirating pump.
Fire 08 00358 g005
Fire 08 00358 g006
Figure 6. Research paths of fire detection technology in large-space environments.
Figure 6. Research paths of fire detection technology in large-space environments.
Fire 08 00358 g006
Fire 08 00358 g007
Figure 7. Recognition process.
Figure 7. Recognition process.
Fire 08 00358 g007
Fire 08 00358 g008
Figure 8. Schematic diagram of indoor image acquisition process. (a) Image acquisition device; (b) setting different acquisition distances; (c) setting different acquisition directions; (d) different occlusion experiment occluders; (e) ignited experimental materials.
Figure 8. Schematic diagram of indoor image acquisition process. (a) Image acquisition device; (b) setting different acquisition distances; (c) setting different acquisition directions; (d) different occlusion experiment occluders; (e) ignited experimental materials.
Fire 08 00358 g008
Fire 08 00358 g009
Figure 9. Grad-CAM visualization of smoke detection results [79].
Figure 9. Grad-CAM visualization of smoke detection results [79].
Fire 08 00358 g009
Fire 08 00358 g010
Figure 10. The detection results of flames using AH-YOLO: (a) original scene image; (b) detection results [84].
Figure 10. The detection results of flames using AH-YOLO: (a) original scene image; (b) detection results [84].
Fire 08 00358 g010
Table 1. Comparison of fire detector technologies: advantages, disadvantages, and performance characteristics.
Table 1. Comparison of fire detector technologies: advantages, disadvantages, and performance characteristics.
Detector
Type		Working
principle		AdvantagesLimitationsTypical
Applications
1. Point-type Smoke/Temperature DetectorsSmoke Sensing: Detects smoke particles via ionization or photoelectric
effects.
Heat Sensing: Triggers upon reaching a fixed temperature or a rapid rate of rise in ambient temperature.		Mature technology with low cost. Simple installation and maintenance. Widely available and versatile. Effective for both flaming and smoldering firesLimited coverage area per unit; requires many devices for large spaces. Height restrictions for installation.
Slow response; relies on smoke diffusion to the sensor. Prone to false alarms from environmental factors (e.g., dust, humidity, airflow).
Unsuitable for large, open, or high-airflow environments.		Enclosed spaces with conventional ceiling heights: offices, hotel rooms, residences, small retail shops.
2. Line-type Beam Smoke DetectorsMeasures the obscuration (light attenuation) of a projected beam between an emitter and a receiver unit to detect smoke.Large area coverage with a single system.
Can be installed at greater heights.
More suitable for large spaces than point-type detectors.		Complex alignment during installation; susceptible to misalignment from vibration. High maintenance requirement; optical lenses must be kept clean. Response delay persists due to reliance on smoke diffusion. Line of sight can be obstructed. Ineffective for non-particulate fires (e.g., clean burning alcohol).Large warehouses, exhibition halls, atriums, sports arenas.
3. Infrared and Ultraviolet Flame DetectorsInfrared Detectors: Sense characteristic infrared radiation patterns from flames.
Ultraviolet Detectors: Sense ultraviolet radiation emitted by flames.		Extremely fast response; does not require smoke propagation. Long detection range. Highly effective for flaming fires.High cost. Susceptible to false alarms from non-fire radiation sources (e.g., welding arcs, sunlight, halogen lamps, heaters). Ineffective for smoldering fires (requires visible flames).
A clear line of sight to the fire is required, yet this line of sight can be obstructed by obstacles.		High-hazard areas with flammable liquids/gases: petrochemical plants, fuel storage facilities, munitions stores.
4. Aspirating Smoke Fire DetectorsActively draws air samples through a network of pipes to a central, highly sensitive laser detection chamber for analysis.Very high sensitivity; can detect invisible combustion particles.
Provides very early warning (incipient stage detection).
Suitable for environments with complex airflows. Flexible pipe network design allows for extensive coverage.		The coverage area of a single device is limited. System complexity results in the highest cost. Demanding installation, design, and commissioning requirements. Requires regular maintenance (filter replacement, airflow checks). Sampling pipes can become clogged with dust or insects.Mission-critical or sensitive sites requiring earliest possible warning: data centers, heritage buildings, high-value archives.
Table 2. Common object recognition algorithms.
Table 2. Common object recognition algorithms.
Two-StageOne-Stage
Rcnn [30,31]YOLO (v8-v12) [32,33,34,35,36,37,38]
Fast rcnn [39]RetinaNet [40]
Faster rcnn [41,42]Mask R-CNN [43]
SPPnet [44]SSD [45,46]
Table 3. Aircraft hangar fire detection: visible light vs. infrared cameras.
Table 3. Aircraft hangar fire detection: visible light vs. infrared cameras.
AspectVisible Light CamerasInfrared Cameras
Detection
Principle		Relies on visible light imaging. Uses algorithms to identify flames based on color, shape and smoke based on texture, and motion characteristics.Detects thermal radiation (infrared energy) emitted by objects. Identifies fire sources through temperature anomalies.
Light
Dependency		Highly dependent on ambient lighting. Performance degrades significantly in darkness, strong backlight, and other complex lighting conditions.Unaffected by visible-light conditions. Operates effectively in total darkness, glare, and any lighting scenario, enabling 24/7 reliable monitoring.
Smoke Detection
Capability		A core strength. Effective at identifying the visual characteristics of smoke. However, performance is limited in dark environments (e.g., hangar corners) where attention becomes dispersed.Cannot directly “see” smoke. Can indirectly detect fire through the heat source or temperature changes caused by smoke.
Flame Detection CapabilityEffective at identifying visible flames. Prone to false alarms from objects with similar shapes and colors.Detects the heat source itself, not its visual manifestation. Effectively identifies hidden fire sources invisible to the naked eye with strong resistance to visual false alarms.
Anti-Interference CapabilityCan perform well in strong sunlight with optimized algorithms, but susceptible to interference from steam, dust, and visual deception.Immune to visual camouflage. Can effectively penetrate smoke, dust, steam, and other particulates, offering significant advantages in low-visibility, harsh environments.
Conclusion for Hangar UseWhile effective under specific conditions, its smoke detection weakness in low light and reliance on visible flames pose reliability risks in the complex, variable hangar environment.Deemed the superior and more reliable solution for hangar applications due to its immunity to light and long-range and direct heat-sensing capabilities, which perfectly match the large, complex environment.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Deng, L.; Wu, S.; Zou, S.; Liu, Q. Large-Space Fire Detection Technology: A Review of Conventional Detector Limitations and Image-Based Target Detection Techniques. Fire 2025, 8, 358. https://doi.org/10.3390/fire8090358

AMA Style

Deng L, Wu S, Zou S, Liu Q. Large-Space Fire Detection Technology: A Review of Conventional Detector Limitations and Image-Based Target Detection Techniques. Fire. 2025; 8(9):358. https://doi.org/10.3390/fire8090358

Chicago/Turabian Style

Deng, Li, Siqi Wu, Shuang Zou, and Quanyi Liu. 2025. "Large-Space Fire Detection Technology: A Review of Conventional Detector Limitations and Image-Based Target Detection Techniques" Fire 8, no. 9: 358. https://doi.org/10.3390/fire8090358

APA Style

Deng, L., Wu, S., Zou, S., & Liu, Q. (2025). Large-Space Fire Detection Technology: A Review of Conventional Detector Limitations and Image-Based Target Detection Techniques. Fire, 8(9), 358. https://doi.org/10.3390/fire8090358

Article Metrics

Back to TopTop