ETRI-Knowledge Sharing Plaform



논문 검색
구분 SCI
연도 ~ 키워드


학술대회 Zero-shot Fire And Arson Detection Using Textual Descriptions
Cited 0 time in scopus Download 3 time Share share facebook twitter linkedin kakaostory
전호범, 김형민, 김도형, 김재홍
International Conference on Information and Communication Technology Convergence (ICTC) 2022, pp.1563-1568
22PS2700, 인간과 로봇의 물리적, 인지적 상호작용을 통하여 정서 교감이 가능한 반려로봇 기술 개발, 김도형
Recently, fire detection methods have leveraged deep learning technology to detect fire and flame texture information in images. However, training-based methods require large amounts of data to detect, and data collecting is expensive due to the fire risk. In addition, the existing dataset extracted from web videos lacks diversity in backgrounds. This paper proposes a method for detecting fire situations based on natural language descriptions to solve the data diversity problem. Our approach compares images and descriptions of the fire situation for detecting fire without training. In addition, the detector is simultaneously reasoning human arson behavior, enabling early fire detection in the surveillance videos. Our fire detection method demonstrates superior or competitive performance without training compared to other training-based methods. Furthermore, the surveillance performance of our text-based fire detector was investigated by datasets including inclement weather and nighttime situations. As a result, we confirm text-based fire detection method has high generalization performance and usability. Our approach could eventually lead to future interpretable anomaly detection studies.
KSP 제안 키워드
Competitive performance, Detection Method, Early fire detection, Fire Risk, Fire detector, Generalization performance, Surveillance performance, Surveillance video, Texture information, Web videos, Zero-shot