A carregar...
Publicação
Deep Learning based self-localization in virtual environments
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.92 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Machine learning systems can greatly improve tasks that are otherwise difficult, one of which is self-localization. This thesis explores the use of Convolutional Neural Networks to achieve selflocalization in 3D virtual environments by developing a supervised learning model based on the VGG CNN architecture. The virtual 3D environments are created using the 3D modeling program Blender, which includes a Python API that enables the scene to be constructed entirely in code. Resorting to this process addresses one of the primary issues common to every machine learning algorithm: the acquisition of reliable datasets that accurately represent the problem’s dimensions, along with accurately labeled items. The easily modifiable environment helps in the study of the impact of the scene on the performance of the localization model. The experiments conducted explored a VGG model that receives as input a pair of images (stereo) that capture opposite views of the scene, as an attempt to illustrate more environmental context. These experiments were also conducted with a single image as input version of the model, to compare the performances. The stereo iteration, while providing acceptable results in some environments, emonstrated to be under-performing compared to the single version in every scene that was tested. The experiments also demonstrated the influence that the datasets have over the model’s performance, where changes to the environments that build these datasets were needed to achieve decent results. Future work that further explores this discrepancy was also presented.
Descrição
Tese de Mestrado, Engenharia Informática, 2025, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Deep Learning Localization, Blender Convolutional Neural Network VGG