ETRI-Knowledge Sharing Plaform

ENGLISH

성과물

논문 검색
구분 SCI
연도 ~ 키워드

상세정보

학술대회 Ensemble Grid Formation to Detect Potential Anomalous Regions Using Context Encoders
Cited 0 time in scopus Download 6 time Share share facebook twitter linkedin kakaostory
저자
무함마드, 마셀라, 이승익, 신호철
발행일
201810
출처
International Conference on Control, Automation and Systems (ICCAS) 2018, pp.661-665
협약과제
18HS4900, 실외 무인 경비 로봇을 위한 멀티모달 지능형 정보분석 기술 개발, 신호철
초록
Unsupervised visual anomaly detection is currently one of the most challenging machine learning problems. Various techniques have been specifically developed for certain limited domains while some are more generic towards a broader scope of applications. This paper aims to investigate a novel solution for general anomaly detection in surveillance videos by modeling patterns and objects that appear normally in the videos and then using this model to detect the anomalous objects by exploiting image reconstruction methodologies. This approach is inspired by the recent progression in development of robust semantic in-painting techniques. For our experiments, Context Encoders are used for the said purpose. Context encoders are proven successful to reconstruct missing holes in images based on the non-hole parts. Our proposed methodology is semi-supervised which means it does not require an annotated dataset however the videos of cases containing normal scenes are required separately to train the system. Various experiments suggest that the proposed methodology can successfully locate potentially anomalous images from the normal ones. This paper discusses in depth the possibilities of adopting such systems for general anomaly detection, pros, cons as well as the limitations of the overall methodology.
키워드
Anomaly Detection, Image Reconstruction, Semantic Inpainting, Surveillance Robot
KSP 제안 키워드
Anomalous objects, Grid Formation, Image reconstruction, In-painting techniques, Semi-supervised, Surveillance robot, Surveillance video, anomaly detection, machine Learning