Detecção e rastreamento de múltiplos objetos utilizando redes profundas no contexto de mapeamento de formigueiros em plantação de Eucaliptos

• Ano de defesa: 2022
• Autor(a) principal: GIAN LUCAS DA SILVA RAMALHO
• Orientador(a): Jonathan de Andrade Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Agricultura de Precisão, detecção de objetos, Deep Learning, Rastreamento de Objetos
• Link de acesso: https://repositorio.ufms.br/handle/123456789/5146

The forestry sector enables economic and environmental development, offering employment and income to the population and helping to reduce climate change. According to IBGE, in 2019, the area of forests cultivated throughout the national territory reached a total of 9.98 million hectares. Eucalyptus cultivation represents approximately 76%, equivalent to 7.61 million hectares. In the forest plantations present in Brazil, one of the main pests that intensely affect production, are leaf-cutting ants. These insects consume a lot of vegetation, attacking different plant species and causing defoliation to death. of the plant, regardless of its size, from seedlings to trees. To fight ants, chemical products are used, along with plantation monitoring. You can apply detection and tracking of objects in images of the plantations, to assist in the monitoring of the plantation and the anthills. The detection and tracking of objects in this study fit the context of tracking multiple objects, Multiple Object Tracking (MOT). The MOT task refers to locating multiple objects, identifying them, and calculating their trajectories. individual images in a sequence of images. In this study, three object detectors, Faster R-CNN, RetinaNet, and VFNet, along with the Tracktor tracking methods, Byte Tracker Deep Sort in addition to the proposal of a method based on the SORT Method, for tracking anthills. Evaluations of object detection and tracking methods were carried out, and the best tracking results obtained were using the RetinaNet detector which achieved 0.817 of Average Precision (AP), 53,004 of Higher Order Tracking Accuracy (HOTA) with method Byte Tracker, 47,120 HOTA with the Proposed Method and 43,426 HOTA with the Deep Sort. Although the Byte Tracker indicates a superior HOTA result, the Method Proposed excels in counting objects, out performing other methods tracking.