Development of synthetic visual datasets to enhance weed and crop perception in agricultural robots

This project is developed within the framework of Earth Rover, a company focused on designing advanced agricultural robots. Specifically, it aims at the development of a robot capable of eliminating weeds by concentrating energy pulses to boil the stem cells located in the meristem of the plants. Th...

ver descrição completa

Detalhes bibliográficos
Autor: Escribano Garcia, Victor
Formato: tesis de maestría
Fecha de publicación:2025
País:España
Recursos:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/442576
Acesso em linha:https://hdl.handle.net/2117/442576
Access Level:acceso abierto
Palavra-chave:Robotics
Artificial intelligence
Computer vision
Robòtica
Intel·ligència artificial
Visió per ordinador
Àrees temàtiques de la UPC::Enginyeria agroalimentària
Descrição
Resumo:This project is developed within the framework of Earth Rover, a company focused on designing advanced agricultural robots. Specifically, it aims at the development of a robot capable of eliminating weeds by concentrating energy pulses to boil the stem cells located in the meristem of the plants. The main objective is to intervene at the earliest growth stage, when weeds are typically small dicotyledons, ensuring a more efficient and precise elimination. Manual data annotation for training computer vision models is complex, costly, and often imprecise, especially when weeds form dense clusters, making precise human annotation nearly impossible. In this context, the use of synthetic datasets emerges as a powerful tool that not only assists in annotating complex scenes but also introduces domain randomization that is difficult to achieve with real-world images (e.g., variations in lighting conditions, terrain types, etc.). This work develops a system that strives to realistically replicate agricultural environments using Blender and its Geometry Nodes system for the procedural generation of crops and weeds. It enables the automatic creation of synthetic images along with segmentation masks, depth maps, and keypoint annotations (focused on the meristem). Synthetic datasets are increasingly established as a key resource to exponentially accelerate the development of computer vision systems and machine learning models in agricultural robotics.