Banca de DEFESA: MATEUS FILIPE SILVA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
DISCENTE : MATEUS FILIPE SILVA
DATA : 17/07/2026
HORA: 13:30
LOCAL: Sala Virtual: meet.google.com/ojd-yxcv-gnb
TÍTULO:

An Artificial Intelligence-Based Approach for the Localization and Topological Navigation of Intelligent Vehicles

PALAVRAS-CHAVES:

Topological localization, Computer vision, Machine learning, Intelligent vehicles, Autonomous navigation.

PÁGINAS: 89
GRANDE ÁREA: Engenharias
ÁREA: Engenharia Elétrica
RESUMO:

Localization is a fundamental stage in navigation systems applied to intelligent vehicles,
as it enables the association of the vehicle’s current position with a structured repre-
sentation of the environment. In urban scenarios, this task is challenging due to the
presence of intersections, lighting variations, changes in perspective, visual ambiguities,
and variations in image acquisition conditions. In this context, this dissertation proposes
and evaluates a topological localization approach based on computer vision and deep
learning, structured into two decision stages. The first stage performs binary classifica-
tion of the road context, distinguishing non-intersection road segments from intersection
regions. The second stage is activated only when an intersection is detected and aims
to identify the corresponding topological node in the map. To support the research, an
infrastructure for acquiring, organizing, and automatically labeling georeferenced images
was developed, using GNSS data and a previously defined topological map. The con-
structed dataset contains 3,132 images acquired in a real urban environment, of which
1,637 correspond to non-intersection road segments and 1,495 to intersection regions. The
MobileNetV3-Small, EfficientNet-B0, and ShuffleNet V2 architectures were evaluated in
both pipeline tasks. The final configuration used MobileNetV3-Small in the first stage
and EfficientNet-B0 in the second stage, achieving a global accuracy of 94.5% on the
test set and an average processing rate of 47.61 FPS. Robustness experiments were also
conducted under variations in brightness, contrast, blur, and rotation, in addition to field
tests on four real urban routes. The results indicated that the pipeline achieved high
performance under controlled validation conditions and tolerated moderate photometric
perturbations, but showed sensitivity to loss of sharpness, camera misalignment, and ad-
verse lighting conditions. In the field tests, the pipeline accuracy ranged from 44.70%
to 74.74%, indicating that generalization to real routes still depends on greater dataset
diversity, improvement of topological node classification, and the future incorporation of
temporal or spatial consistency mechanisms. Thus, this dissertation demonstrates the ex-
perimental feasibility of visual topological localization for intelligent vehicles, while also
delimiting its main operational limitations.

MEMBROS DA BANCA:
Presidente - 2362923 - GIOVANI BERNARDES VITOR
Interno - ***.230.456-** - LUIZ CARLOS BAMBIRRA TORRES - UFOP
Interno - 1183761 - SANDRO CARVALHO IZIDORO
Externo à Instituição - JANITO VAQUEIRO FERREIRA - UNICAMP