직하방 안전 감시 시스템을 위한 자기 주의 증류 기반 3차원 회전 객체 인식 방법 (Self Attention Distillation-based Rotational 3D Object Recognition for Nadir FOV Safety Surveillance System)

서지정보

· 발행기관 : 제어·로봇·시스템학회
· 수록지 정보 : 제어.로봇.시스템학회 논문지 / 30권 / 12호 / 1422 ~ 1429페이지
· 저자명 : 우민우, 이충환, 김병학

초록

Accidents originating from mobile cranes account for approximately 13% of all accidents in the construction industry, which is significantly higher than all other accident cases. Thus, there has been a strong demand for technology solutions to prevent collisions between salvage and surrounding objects. We herein propose a safety surveillance system using rotational 3D object recognition in the nadir field of view (vertically downward FOV) based on self-attention distillation (SAD). We developed training, validation, and test datasets for the environment of an actual construction site to develop a rotational object detection model for the proposed system. Moreover, we introduced an SAD method for the backbone network to improve the representation of the backbone network and guarantee the accurate detection of objects and salvages. Overall, the proposed rotational object detection model, which is based on Real-Time Models for object Detection (RTMDet) and SAD, could achieve a performance of over 78% mean average precision (mAP).

영어초록

Accidents originating from mobile cranes account for approximately 13% of all accidents in the construction industry, which is significantly higher than all other accident cases. Thus, there has been a strong demand for technology solutions to prevent collisions between salvage and surrounding objects. We herein propose a safety surveillance system using rotational 3D object recognition in the nadir field of view (vertically downward FOV) based on self-attention distillation (SAD). We developed training, validation, and test datasets for the environment of an actual construction site to develop a rotational object detection model for the proposed system. Moreover, we introduced an SAD method for the backbone network to improve the representation of the backbone network and guarantee the accurate detection of objects and salvages. Overall, the proposed rotational object detection model, which is based on Real-Time Models for object Detection (RTMDet) and SAD, could achieve a performance of over 78% mean average precision (mAP).

참고자료

· 없음