에지 디바이스를 활용한 실시간 말벌 모니터링 시스템의 설계 및 구현 (Design and Implementation of Real-time Wasp Monitoring System using Edge Device)

서지정보

· 발행기관 : 한국멀티미디어학회
· 수록지 정보 : 멀티미디어학회논문지 / 25권 / 12호 / 1826 ~ 1839페이지
· 저자명 : 정윤주, 황호석, 권영재, 이철희

초록

In this paper, an edge-enabled wasps monitoring system based on state of the art deep learning model is proposed to classify wasps species and transmit their information to clients for pest control and ecology research. We utilized and evaluated the lightweight models of YOLOv5, YOLOX, and YOLOv7 series to select the optimal object detection model for wasps. In order to improve the recognition performance of small objects such as wasps, a new tiling method is proposed which can reclassify objects located on boundary area between tiles. Our monitoring system is composed by local GPU-based edge computers, Server, and Web-based reporting system. In our experiment, the YOLOX- nano model showed the highest performance among YOLO-based models above in terms of recall, precision, Fi-score, and mAP@IOUs for our custom test data set. However, even the YOLOX-nano model also requires the tiling process for a better bounding box generation for small objects under the real environment. As a result, the proposed wasps monitoring system showed meaningful efficiency for applying in the field of pest control and ecology research.

영어초록

In this paper, an edge-enabled wasps monitoring system based on state of the art deep learning model is proposed to classify wasps species and transmit their information to clients for pest control and ecology research. We utilized and evaluated the lightweight models of YOLOv5, YOLOX, and YOLOv7 series to select the optimal object detection model for wasps. In order to improve the recognition performance of small objects such as wasps, a new tiling method is proposed which can reclassify objects located on boundary area between tiles. Our monitoring system is composed by local GPU-based edge computers, Server, and Web-based reporting system. In our experiment, the YOLOX- nano model showed the highest performance among YOLO-based models above in terms of recall, precision, Fi-score, and mAP@IOUs for our custom test data set. However, even the YOLOX-nano model also requires the tiling process for a better bounding box generation for small objects under the real environment. As a result, the proposed wasps monitoring system showed meaningful efficiency for applying in the field of pest control and ecology research.

참고자료

· 없음