The growing use of unmanned aerial vehicles (UAVs) equipped with RGB optical sen-
sors, thermal cameras, and LiDAR has expanded applications in structural inspection,
topographic surveying, and environmental monitoring. Reconstructions based on a single
modality, however, still face limitations in accuracy and structural fidelity.
This thesis proposes the hybrid architecture GaussianFusion, which performs early
multimodal integration, unsupervised structural clustering, and continuous rendering via
Gaussian-based splatting. Although the name includes “Fusion,” the proposal adopts a
modular integration approach, enabling visual, thermal, and depth information to interact
from the outset in a lightweight and extensible manner. This modularity facilitates the
incorporation of new sensors and applications in robotic systems.
Validation was carried out in three real urban scenarios: CTEx, Glória, and Quinta da
Boa Vista. Comparisons among point clouds from different sensor combinations were
performed using geometric metrics (RMSE, Chamfer, Hausdorff) and clustering indices
(Silhouette, Davies-Bouldin). RGB-LiDAR integrations achieved the lowest geometric
deviations; thermal data increased interpretability without compromising accuracy. The
resulting clusters were structural, reflecting properties such as elevation, density, and
emissivity.
On the Glória dataset, Gaussian-based splatting was evaluated for continuous rendering
of the church façade. The tests demonstrated photorealistic visualization and structural
continuity, but also revealed limitations when angular density was insufficient, highlighting
the method’s sensitivity to viewpoint distribution.
A benchmark study was also conducted with the Point-BERT model, comparing latent
projections, clustering metrics, and computational costs. Results show that the hybrid
system achieves competitive performance in unsupervised clustering, balancing structural
quality and operational efficiency.
The main scientific contribution lies in demonstrating that early multimodal integration,
combined with unsupervised structural clustering and continuous rendering, provides
geometric and radiometric consistency in complex environments. The proposed architecture
is applicable to civil engineering, precision agriculture, climate impact assessment, and
security and defense, showing consistency, scalability, and generalization capability for 3D
modeling in real environments.