Improving Global Localization Algorithms for Mars Rovers with Neural Networks
Author
Moreno I Caireta, Iñigo
Term
4. semester
Education
Publication year
2021
Submitted on
2021-06-03
Pages
38
Abstract
Denne afhandling undersøger, om et siamesisk neuralt netværk—et par identiske netværk, der lærer at afgøre, om to input ligner hinanden—kan understøtte global lokalisering for Mars-rovere. Afhandlingen beskriver hele forløbet: indsamling og behandling af to datasæt, opbygning af modellen og justering af dens læringsindstillinger (hyperparametre). Derefter testes modellen, og dens ressourceforbrug måles: den fylder op til 9 megabyte og giver forudsigelser på 18 millisekunder. Til sidst vises, hvordan modellen kan bruges i en algoritme til global lokalisering ved at implementere en glidende vindue-tilgang, der gennemgår en sekvens trin for trin. Denne tilgang giver lovende resultater og ser ud til at klare sig bedre end en tidligere løsning.
This thesis explores whether a Siamese Neural Network—a pair of identical networks trained to judge whether two inputs are alike—can support global localization for Mars rovers. It details the end-to-end development: acquiring and processing two datasets, building the model, and tuning its learning settings (hyperparameters). The model is then tested and its efficiency measured: it occupies up to 9 megabytes and produces predictions in 18 milliseconds. Finally, the thesis shows how to apply the model within a global localization algorithm by implementing a sliding-window approach that scans a sequence step by step. This approach delivers promising results and appears to perform better than a previous solution.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents