Full text
| |
| Author(s): |
Guilherme Vicentim Nardari
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | São Carlos. |
| Institution: | Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB) |
| Defense date: | 2023-02-06 |
| Examining board members: |
Roseli Aparecida Francelin Romero;
Valdir Grassi Junior;
Josue Junior Guimarães Ramos;
Denis Fernando Wolf
|
| Advisor: | Roseli Aparecida Francelin Romero |
| Abstract | |
While overhead data can provide general information about a forest, inside the forest, we can identify understory plants and measure the diameter and count of the trunks of each tree. Currently, specialists rely on human expeditions to get these measurements, which can be slow, expensive, and dangerous. For this reason, robots that can autonomously navigate and extract data from inside the forest could revolutionize how we monitor forests worldwide and the amount of information we have about them. In forestry, the lack of reliable GPS signal, uneven terrain covered by plants and leaves, and trees with branches moving with the wind are a few of the challenges posed. These factors can create shortcomings for classic algorithms as some assumptions may not be valid in this environment. Semantic information, such as classes and forms of objects expected in the environment is a promising way to increase the robustness and performance of autonomous systems. In this context, this thesis introduces a framework that uses 3D data provided by LiDAR or stereo cameras to identify semantic information using neural networks. This information is used to identify trees and model them as cylinders, creating a semantic map. Our formulation allows the incorporation of noisy estimates that can be refined with the arrival of new sensor readings and external measurements to increase the frameworks robustness. Using the semantic map generated by our framework, we propose an algorithm capable of generating unique forest location descriptors that are visually highly similar. These descriptors can be used to identify previously visited locations and feedback to reduce accumulated errors in location estimates. We present several experiments in simulated environments and real-world Pine forests, demonstrating that our method generates semantic maps that improve the quality of the robots location estimates and generate informative maps with information on the individual count and the trunk diameter of each tree. Furthermore, the semantic representation of the data obtained by the sensors is much more computationally efficient, as it summarizes the raw data in a semantic geometric model with few parameters. (AU) | |
| FAPESP's process: | 17/17444-0 - Plantation monitoring using heterogeneous robots |
| Grantee: | Guilherme Vicentim Nardari |
| Support Opportunities: | Scholarships in Brazil - Doctorate (Direct) |