Spatial Computing Frameworks for Adaptive Disaster Response and Visual Place Recognition

In the immediate aftermath of a disaster, acquiring information regarding the status of surviving infrastructure is imperative. However, data collected during this period often exhibits inconsistencies, omissions, and errors. We have designed multiple systems to proficiently validate and amalgamate crowd-sourced paths during disasters. One is DeimosBC, a novel post-disaster crowd-sourcing system which relies on a blockchain to provide robustness and decentralization, and the ability for multiple disparate users to contribute their effort to a collaborative task. The other is Proteus, which generates a traversable map from volunteer collected data in a post-disaster scenario. Given a set of collected GPS trajectories per volunteer Proteus can combine them to obtain a single set of connected edges that represent the ground truth quickly and accurately, independent of how the data is collected. Precise geolocation on the Lunar surface is crucial for future exploration, scientific research, and potential human settlement on the Moon. Unlike Earth, the Moon lacks a global positioning system (GPS) infrastructure, presenting significant challenges for accurate positioning and navigation. One potential solution to provide geolocation on the Moon is Visual Place Recognition (VPR), which involves matching one or multiple image sets in order to determine which images show the same places in the world. As there is a severe lack of lunar imagery to train our models on, we have created a model, Despina, for synthesis of high-fidelity location-specific and elevation-realistic Lunar horizon imagery using generative models and digital elevation models. We adapted existing traditional VPR algorithms to account for the unique visual characteristics of the lunar surface, to create a model called Galatea. Galatea uses modified Sequential Delta Descriptors to ignore irrelevant parts of the image (i.e. the common background), while effectively capturing and describing important parts and their relationship within the scene (e.g. how many rocks or craters and where they are in the scene). For feature extraction, we fine-tune the state-of-the-art DINOv2 model with Apollo imagery, which leads to fast convergence and gives us stronger predictive power for our descriptors. Finally, we use the Optimal Transport approach for image retrieval/matching as well as feature matching.

Name thesis:degree_name

Doctor of Philosophy

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Houston

Year dc:date.issued

2025

Author dc:creator

• Sarbajna, Raunak 1990-

Advisor dc:contributor.advisor

• Eick, Christoph F.

Committee members dc:contributor.committeemember

• Tsekos, Nikolaos V.
• Mang, Andreas
• Pavlidis, Ioannis T.

• Computer science

Language dc:language.iso

English

Handle dc:identifier.uri

https://hdl.handle.net/10657/20688

OAI identifier oai:identifier

oai:uh-ir.tdl.org:10657/20688