Arlotta, Andrea (2025) Smart vehicles for active monitoring in Precision Agriculture settings. [Tesi di dottorato]

Documento PDF THESIS_ARLOTTA_ANDREA_PHDAI_37o.pdf Visibile a [TBR] Amministratori dell'archivio Download (31MB) | Richiedi una copia

Abstract

Effective Precision Agriculture demands efficient methods for monitoring crop status, often requiring autonomous robots to operate in complex field conditions. This thesis focuses on developing intelligent control strategies for active monitoring: while robust perception is essential, the core challenge addressed is enabling robots to perform active informative path planning that prospectively reasons about how perceptual conditions will affect the outcome of its sensing pipeline. The central focus is on planning trajectories that actively maximize information gain by accounting for how factors such as viewpoint, and environmental variables impact the quality and content of the data gathered by the perception system. The primary contribution is a novel Informative Path Planning framework utilizing Neural Network-driven Model Predictive Control (Neural IPP MPC). By embedding neural network-based perception within the MPC framework, we aim to enhance the robot's ability to perform tasks such as weed detection and monitoring with improved accuracy and efficiency. This allows the robot to anticipate and optimize for information-rich viewpoints without executing the full perception stack during planning. The thesis first establishes a robust ROS-based object detection and EKF tracking framework using RGB-D data as a foundation. The Neural IPP MPC controller is then developed for single-agent scenarios and extended to coordinated multi-agent monitoring with dynamic task allocation and information consensus. Validations via extensive simulations show that the Neural IPP MPC significantly outperforms traditional coverage and reactive strategies in terms of information gathering efficiency and scales effectively to larger environments and collaborative teams. This work demonstrates a viable approach to integrating complex perception awareness into predictive control for efficient, autonomous agricultural monitoring.

Downloads

Actions (login required)

Modifica documento