The thesis addresses an applied mathematics problem at the intersec- tion of computer vision, image processing, and robotics: the visual local- ization of a mobile platform in environments where the Global Positioning System (GPS) is degraded or unavailable, under the explicit resource con- straints of onboard embedded hardware. Two research lines shape the practical setting: the extraction and geometric verification of visual cor- respondences on the vision side, and the recursive fusion of noisy sensor measurements on the state-estimation side; the two have evolved largely independently. The thesis treats them as a single problem of recursive state estimation for a planar mobile platform, into which the quality of the visual matching enters as an adaptation signal that scales the measure- ment covariance. Feature detection, matching, and state estimation are required to run on CPU-only platforms, so the accuracy of every method is quantified against its computational cost rather than in isolation.
The objective of the thesis is to develop a sensor-fusion framework for the localization of a planar mobile robot that combines wheel odometry and GPS, to extend it with a visual position observation derived from satellite image matching whose measurement covariance is scaled by the matching- quality signal of the geometric-verification stage, and to support this with a comparative study of classical and learned local feature descriptors under explicit resource constraints.
Topic: Семінар відділу математичних проблем механіки та теорії керування
Time: Jun 12, 2026 12:00 Kyiv
Join Zoom Meeting
https://us02web.zoom.us/j/89531631507?pwd=Lx3pYwfWlgkAzkqbiZHAlAQbua8SK9.1