MMAsia2024_Fire and Smoke Detection with Burning Intensity Representation_He3DB.pdf

迹部景吾

8页

0次

2025-09-29

免费下载

Fire and Smoke Detection with Burning Intensity Representation

Xiaoyi Han

School of Software Technology,

Zhejiang University

Yanfei Wu

China Mobile (Suzhou) Software

Technology Co., Ltd.

Nan Pu

University of Trento

Zunlei Feng

School of Software Technology,

Zhejiang University

Qifei Zhang

School of Software Technology,

Zhejiang University

Yijun Bei

School of Software Technology,

Zhejiang University

Lechao Cheng

∗

chenglc@hfut.edu.cn

Hefei University of Technology

ABSTRACT

An eective Fire and Smoke Detection (FSD) and analysis system

is of paramount importance due to the destructive potential of re

disasters. However, many existing FSD methods directly employ

generic object detection techniques without considering the trans-

parency of re and smoke, which leads to imprecise localization

and reduces detection performance. To address this issue, a new At-

tentive Fire and Smoke Detection Model (a-FSDM) is proposed. This

model not only retains the robust feature extraction and fusion capa-

bilities of conventional detection algorithms but also redesigns the

detection head specically for transparent targets in FSD, termed

the Attentive Transparency Detection Head (ATDH). In addition,

Burning Intensity (BI) is introduced as a pivotal feature for re-

related downstream risk assessments in traditional FSD method-

ologies. Extensive experiments on multiple FSD datasets showcase

the eectiveness and versatility of the proposed FSD model. The

project is available at https://xiaoyihan6.github.io/FSD/.

CCS CONCEPTS

• Information systems;

KEYWORDS

Fire and Smoke Detection, Attentive Transparency Detection Head,

Burning Intensity

ACM Reference Format:

Xiaoyi Han, Yanfei Wu, Nan Pu, Zunlei Feng, Qifei Zhang, Yijun Bei, and Lechao

Cheng. 2024. Fire and Smoke Detection with Burning Intensity Representa-

tion. In ACM Multimedia Asia (MMASIA ’24), December 3–6, 2024, Auckland,

New Zealand. ACM, New York, NY, USA, 8 pages. https://doi.org/10.1145/

3696409.3700165

Code is available: https://github.com/XiaoyiHan6/FSDmethod.

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than the

author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or

republish, to post on servers or to redistribute to lists, requires prior specic permission

and/or a fee. Request permissions from permissions@acm.org.

MMASIA ’24, December 3–6, 2024, Auckland, New Zealand

ACM ISBN 979-8-4007-1273-9/24/12... $15.00

https://doi.org/10.1145/3696409.3700165

1 INTRODUCTION

Fire, a formidable and highly perilous disaster, can result in sub-

stantial loss of life, property, societal upheaval, and adverse ecolog-

ical consequences across diverse scenes, such as urban conagra-

tions [

], forest res [

], industrial infernos [

], and transporta-

tion incidents [

]. Considering the profound destructive poten-

tial of re disasters, investigating and developing automated Fire

and Smoke Detection (FSD) and analysis systems is of signicant

importance and is urgently needed [

]. Furthermore,

approaches that employ generic object detection methods in FSD

without taking into account the transparency of re and smoke

undermine the overall performance of FSD methods and may even

result in instances of missed detection and false alarms. Addition-

ally, generic object detection algorithms currently lack an eective

linking mechanism for seamless implementation in downstream

wildre risk assessment tasks.

It is found that transparent ames and smoke frequently appear

during re disasters due to the distinctive features of combustible

materials, the environmental conditions of the re, and the unique

optical properties of certain substances. This transparency eect

can cause indistinct boundaries between the ame or smoke tar-

gets and the background in FSD scenarios, ultimately degrading

the performance of FSD. In addition, the phenomenon may further

lead to missed or false detection of ames or smoke, posing a chal-

lenge to the accuracy of FSD. To mitigate these drawbacks, unlike

existing methods [

] that directly apply generic object detection

algorithms to FSD, a new Attentive Transparency Detection Head

(ATDH) is proposed, which is designed for adaptive enhancement

of transparency features across multiple scales and channels. ATDH

is compatible with most generic object detection algorithms to en-

hance the accuracy of FSD.

Moreover, a Relative Burning Intensity (BI) is dened to repre-

sent BI. To the best of our knowledge, this is the rst representation

of BI. This design serves as an indicator to measure the severity

of combustion and endows generic FSD algorithms with a novel

and essential ability for re-related risk assessment, which plays a

crucial role in determining the allocation of reghting resources,

developing eective re management strategies, and improving the

eciency of re suppression eorts.

In summary, the main contributions are as follows: 1) A novel

arXiv:2410.16642v1 [cs.CV] 22 Oct 2024

MMASIA ’24, December 3–6, 2024, Auckland, New Zealand Xiaoyi Han. et al.

Attentive Fire and Smoke Dete ction Model (a-FSDM) is pro-

posed. Retaining the advantages of generic detection algorithms

in feature extraction and fusion, the re-related target detection

head network is redesigned to meet the distinctive needs of FSD.

2) A new Attentive Transparency Detection Head (ATDH) is

presented, which is tailor-made to address the challenge of detect-

ing transparent re and smoke targets by enhancing the distinct

features of transparent targets while suppressing non-target feature

maps. 3) Burning Intensity (BI) is introduced as an indicator to

measure the severity of combustion, and it serves as a key feature

for subsequent downstream assessment of re damage.

2 RELATED WORK

Fire and Smoke Detection (FSD) technology can be categorized into

two types: smoke detection and ame (re) detection. Smoke detec-

tion, which utilizes the characteristics of smoke, typically allows for

earlier detection, thereby minimizing resource and economic losses.

However, it may face challenges in environments where smoke

disperses slowly or is not visible. Conversely, ame detection relies

on the distinct features of ames, oering more easily detectable

signals but often occurring at a later stage, when the re is already

more developed [45].

In terms of FSD task classication, the model described in [

]

involves training pre-trained VGG16 [

] and ResNet50 [

] mod-

els using a custom FSD dataset [

] to enhance FSD performance.

Afterwards, a novel lightweight classication network designed for

re incidents is proposed by FireNet [

]. Subsequently, FireNet-v2,

an improved version of FireNet, achieves a percision of 94.95 [

Moreover, a recent study [

] introduces a deep learning network

that combines CNN and RNN [

] for the purpose of forest re clas-

sication. Subsequently, a novel smoke recognition method based

on dark-channel assisted mixed attention [

], is proposed. How-

ever, the traditional classied FSD task is limited to determining the

existence of re, which cannot provide more valuable information,

such as re location, for reghters.

In recent years, various detection algorithms are incorporated

with the aim of enhancing FSD task systems. For example, Faster

RCNN [

] is introduced in [

]. To reduce false and missed de-

tections, additional algorithms such as SSD [

], R-FCN [

], and

YOLOv3 [

] are integrated by researchers [

], resulting in high

levels of accuracy in their respective datasets. Li et al. [

] demon-

strate the successful application of the widely-used DERT [

] to

FSD tasks. Meanwhile, the FSD network GLCT [

], which merges

CNN and Transformers [

], achieves an mAP of 80.71 in the de-

tection of early re in surveillance video images. However, despite

their eectiveness, Transformers have higher computational re-

source requirements and slower inference times than traditional

CNNs. Subsequently, Venâncio et al. [

] introduce YOLOv5 [

] to

FSD tasks, reporting improved accuracy with a smoke detection

accuracy of 85.88 and a ame detection accuracy of 72.32 on their

surveillance video database. It is noteworthy that traditional object

detection-based FSD models are constructed upon general object

detection algorithms, rather than being specically designed for

the FSD task.

Deep learning techniques are applied to image segmentation [

]

and scene understanding [

] in FSD tasks. Guan et al. [

] develop

MaskSU RCNN, a forest re instance segmentation approach based

on the MS RCNN model. Meanwhile, Perrolas et al. [

] propose

a quad-tree search-based method for localizing and segmenting

res at dierent scales. It is noteworthy that re and smoke in

early-stage FSD present considerable diculties due to their trans-

parency and lack of specicity. The complex and variable nature

of re and smoke results in suboptimal performance of traditional

semantic segmentation techniques [

]. Moreover, in comparison

to semantic segmentation models that require substantial compu-

tational resources for training and inference, detection algorithms

are more appropriate for FSD tasks.

Despite advancements in FSD tasks using object detection meth-

ods, computer vision-based FSD algorithms still lag behind generic

computer vision algorithms. For example, the detection of transpar-

ent foregrounds has rarely been addressed in previous FSD studies.

Moreover, generic object detection does not adequately address

subsequent re-related concerns, including BI, which is crucial for

evaluating the cost of re damage and determining the human re-

sources required to respond to a re.

In order to address the issue of the transparent foreground, a

novel FSD method, namely the Attentive Fire and Smoke Detection

Model (a-FSDM), is presented. The proposed method preserves the

strengths of traditional detection algorithms in feature extraction

and fusion while redesigning the target detection head network

specically for FSD, termed the Attentive Transparency Detection

Head (ATDH). Furthermore, it is assumed that there is a positive

correlation between the severity of re and smoke disasters and

Burning intensity (BI). This indicates that higher BI levels corre-

spond to greater impacts caused by these disasters. To this end, a

novel representation of BI is developed with the aim of facilitat-

ing evaluation for subsequent downstream tasks related to FSD.

Additionally, the proposed algorithm is compared with multiple

multi-scale object detection baselines on various FSD datasets.

3 METHOD

The proposed method contains two parts: the Attentive Fire Smoke

Detection Model (a-FSDM) and the Representation of Burning In-

tensity (BI). In Fig. 1, the a-FSDM is presented, and it consists of

three main components: Feature Extraction, Feature Fusion Group-

ing, and the Attentive Transparency Detection Head (ATDH) for

the FSD task. Feature Fusion Grouping is used to fuse semantic and

spatial information from various convolutional layers to prevent

information loss. The ATDH is responsible for classication, regres-

sion, and centerness.

3.1 Attention Transparency Detection Head

To enhance the model’s comprehension and detection of transpar-

ent ame or smoke targets, a new Attentive Transparency Detection

Head (ATDH) is proposed. The feature map output from the Neck

undergoes four convolutional layers, followed by Global Average

Pooling (GAP) and Max Pooling (MP). GAP ensures that higher

values contribute more to the overall training, while MP assigns

importance to the maximum value [

]. To avoid downplaying the