Seguimiento de plantaciones de café a través de fotogrametría UAS y técnicas de aprendizaje profundo
Loading...
Date
Authors
López-Rojas, Santiago Andrés
Journal Title
Journal ISSN
Volume Title
Publisher
Instituto Tecnológico de Costa Rica
Abstract
El documento presenta una estrategia de detección y conteo de cafetos en ortomosaicos
RGB tomados con drones a baja altura. El trabajo se realizó en el laboratorio
de fotogrametría del Instituto Tecnológico de Costa Rica (ITCR), en colaboración del
Instituto del Café de Costa Rica (Icafe). La red neuronal propuesta se basa en la arquitectura YOLO, y los resultados son comparados con el estado del arte de detección
YOLOv4. Se describe el proceso de confección del conjunto de datos y el preprocesamiento al que se somete, donde todas las imágenes poseen una resolución espacial de alrededor de 2cm/píxel. Se obtiene como resultado del entrenamiento una detección del 92 %, que hace referencia a la exhaustividad para un IoU de 0.5 y se alcanza un mAP@50 de 83.18 %.
The document presents a strategy for the detection and counting of coffee trees in RGB orthomosaics taken with drones at low altitude. The work was carried out in the photogrammetry laboratory of the Technological Institute of Costa Rica (ITCR), in collaboration with the Coffee Institute of Costa Rica (Icafe). The neural network proposal is based on the YOLO architecture, and the results are compared with the state of the art of YOLOv4 detection. The process of making the data set and the preprocessing to which it is subjected is described, where all the images have a spatial resolution of around 2cm/pixel. As a result of the training, a detection of 92% is obtained, which refers to the recall for an IoU of 0.5 and a mAP@50 of 83.18 %.
The document presents a strategy for the detection and counting of coffee trees in RGB orthomosaics taken with drones at low altitude. The work was carried out in the photogrammetry laboratory of the Technological Institute of Costa Rica (ITCR), in collaboration with the Coffee Institute of Costa Rica (Icafe). The neural network proposal is based on the YOLO architecture, and the results are compared with the state of the art of YOLOv4 detection. The process of making the data set and the preprocessing to which it is subjected is described, where all the images have a spatial resolution of around 2cm/pixel. As a result of the training, a detection of 92% is obtained, which refers to the recall for an IoU of 0.5 and a mAP@50 of 83.18 %.
Description
Proyecto de Graduación (Maestría en Electrónica) Instituto Tecnológico de Costa Rica, Escuela de Ingeniería Electrónica, 2022.
Citation
Collections
Endorsement
Review
Supplemented By
Referenced By
Creative Commons license
Except where otherwised noted, this item's license is described as acceso abierto
